<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Jeffrey Turov</title>
    <description>The latest articles on DEV Community by Jeffrey Turov (@jeffreyturov).</description>
    <link>https://dev.to/jeffreyturov</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4050944%2Fabfb8306-248c-4a94-8471-d954f41366c3.png</url>
      <title>DEV Community: Jeffrey Turov</title>
      <link>https://dev.to/jeffreyturov</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/jeffreyturov"/>
    <language>en</language>
    <item>
      <title>I published 30 runnable API examples — here is what the real outputs taught me</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Sun, 08 Nov 2026 14:31:57 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/i-published-30-runnable-api-examples-here-is-what-the-real-outputs-taught-me-2f6j</link>
      <guid>https://dev.to/jeffreyturov/i-published-30-runnable-api-examples-here-is-what-the-real-outputs-taught-me-2f6j</guid>
      <description>&lt;p&gt;A green run status is not enough to trust an API. I want to inspect the dataset, understand the billing mode and see what happened when a provider or a source misbehaved.&lt;/p&gt;

&lt;p&gt;I maintain 15 public Apify Actors at Adaga solutions. The toolbox now links 30 prefilled public tasks covering local-business data, reviews, social profiles, RAG, AI-brand evidence, hotel rate cards, French fuel prices and OpenAPI monitoring.&lt;/p&gt;

&lt;p&gt;This is not a claim of 30 customers or 30 universally comparable benchmarks. It is a set of small, reproducible inputs with their outputs and limitations.&lt;/p&gt;

&lt;p&gt;Here are five lessons from the recorded runs.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. A baseline is not a change-detection result
&lt;/h3&gt;

&lt;p&gt;The Stripe OpenAPI example (&lt;code&gt;jW0uxs1wQyafDZjEX&lt;/code&gt;) fetched the public Stripe specification and stored a baseline. Its output says exactly that future runs will diff against this version.&lt;/p&gt;

&lt;p&gt;It does not prove that Stripe has no breaking changes. That requires a later comparison against a changed specification.&lt;/p&gt;

&lt;p&gt;Try the bounded task:&lt;br&gt;
&lt;a href="https://apify.com/travelmonitorlab/api-breaking-change-radar/examples/stripe-openapi-breaking-change-watch" rel="noopener noreferrer"&gt;https://apify.com/travelmonitorlab/api-breaking-change-radar/examples/stripe-openapi-breaking-change-watch&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  2. AI answers need model, prompt and error provenance
&lt;/h3&gt;

&lt;p&gt;A fixed-model Yasha Hotel run (&lt;code&gt;rTzuzBVg0h45cFz3L&lt;/code&gt;) measured three answers from &lt;code&gt;recursive/flint-v1.0:free&lt;/code&gt;. All three mentioned the hotel. None was marked as citing the official &lt;code&gt;yashahotel.com&lt;/code&gt; domain.&lt;/p&gt;

&lt;p&gt;Two answers contained one model-provided citation each. That is why total citation URLs and official-domain citations are separate fields.&lt;/p&gt;

&lt;p&gt;A second proof (&lt;code&gt;dk1Hi70ELwgSUp01I&lt;/code&gt;) compared Adaga solutions and Manus across prompts and models. It preserved three real answers across two models and one provider error. The failed provider call was not counted as a successful answer.&lt;/p&gt;

&lt;p&gt;Public dry-run configuration:&lt;br&gt;
&lt;a href="https://apify.com/travelmonitorlab/ai-brand-visibility-tracker/examples/brand-p2-multi-model-competitors" rel="noopener noreferrer"&gt;https://apify.com/travelmonitorlab/ai-brand-visibility-tracker/examples/brand-p2-multi-model-competitors&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;These are observations for named models, prompts and timestamps — not universal AI-visibility scores.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Preserve source anomalies instead of making the data look tidy
&lt;/h3&gt;

&lt;p&gt;Hotel Rate Monitoring run &lt;code&gt;7ObEuoyCma8msaGhq&lt;/code&gt; fetched Hotel Martha and Hostellerie Val Fleuri from Visit Luxembourg in 3.686 seconds. It returned two cards with zero source errors.&lt;/p&gt;

&lt;p&gt;One source labels a suite with minimum 201 and maximum 195. The Actor keeps those values and flags the anomaly rather than swapping them into a plausible-looking range.&lt;/p&gt;

&lt;p&gt;Every card includes the source URL, fetch timestamp, HTTP status and HTML SHA-256 hash. The values are published min/max rates, not live availability, date-specific quotes or OTA parity scores.&lt;/p&gt;

&lt;p&gt;Real dataset:&lt;br&gt;
&lt;a href="https://api.apify.com/v2/datasets/HzYByHb18ia1v9p0T/items" rel="noopener noreferrer"&gt;https://api.apify.com/v2/datasets/HzYByHb18ia1v9p0T/items&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Public task:&lt;br&gt;
&lt;a href="https://apify.com/travelmonitorlab/travel-monitor-launch/examples/mersch-luxembourg-published-hotel-rates" rel="noopener noreferrer"&gt;https://apify.com/travelmonitorlab/travel-monitor-launch/examples/mersch-luxembourg-published-hotel-rates&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Validate the resolved business before using review data
&lt;/h3&gt;

&lt;p&gt;A Luxembourg Review Radar run requested &lt;code&gt;Le Bouquet Garni Luxembourg&lt;/code&gt; and returned five real review records. The output's resolved place was &lt;code&gt;Restaurant Auberge Gourmande&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;That mismatch is useful evidence: a workflow must preserve both query and resolved place, and a human should validate the mapping before assigning reviews to a prospect or sending an alert.&lt;/p&gt;

&lt;p&gt;A related five-review report for Esch-sur-Alzette returned &lt;code&gt;clean&lt;/code&gt; because none of those five recent reviews was at three stars or below. That says nothing about reviews outside the bounded sample.&lt;/p&gt;

&lt;p&gt;Public task:&lt;br&gt;
&lt;a href="https://apify.com/travelmonitorlab/review-radar/examples/luxembourg-city-restaurant-review-triage" rel="noopener noreferrer"&gt;https://apify.com/travelmonitorlab/review-radar/examples/luxembourg-city-restaurant-review-triage&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  5. Small outputs are enough to test the contract
&lt;/h3&gt;

&lt;p&gt;A Lyon Maps run returned three restaurants in 45.305 seconds: Carmelo, ART Restaurant Lyon and Maison Moly. That is enough to validate fields and cost before scaling, but it is not a complete market list.&lt;/p&gt;

&lt;p&gt;The French fuel example returned 226 station records after filtering Gazole data to departments 69 and 75. The sample station and price are a timestamped snapshot, not a claim about today's cheapest fuel.&lt;/p&gt;

&lt;p&gt;The Bubbles Coiffure audit analyzed three reviews and recorded a presence score of 70 in dry-run billing. It is a bounded audit result, not a scientific ranking.&lt;/p&gt;

&lt;h3&gt;
  
  
  What the 30 public tasks are for
&lt;/h3&gt;

&lt;p&gt;Start with the smallest input, inspect the actual dataset, then decide whether the tool fits your workflow. Technical dry-runs are labelled as such; they skip Actor event charges but not necessarily compute, browser, proxy or external-provider usage.&lt;/p&gt;

&lt;p&gt;Catalogue and proofs:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://adagasolutions.lu/toolbox/" rel="noopener noreferrer"&gt;https://adagasolutions.lu/toolbox/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://adagasolutions.lu/apis/" rel="noopener noreferrer"&gt;https://adagasolutions.lu/apis/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Source and integration examples:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/jeffreyturov-dev/apify-scraping-toolbox" rel="noopener noreferrer"&gt;https://github.com/jeffreyturov-dev/apify-scraping-toolbox&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Which acceptance artifact do you inspect first when evaluating an API: the input, one dataset item, the source provenance or the billing events?&lt;/p&gt;

</description>
      <category>apify</category>
      <category>webscraping</category>
      <category>mcp</category>
      <category>ai</category>
    </item>
    <item>
      <title>Aura — A Consent-First Garden That Blooms With Your Real Pulse</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Fri, 06 Nov 2026 08:50:40 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/aura-a-consent-first-garden-that-blooms-with-your-real-pulse-24hi</link>
      <guid>https://dev.to/jeffreyturov/aura-a-consent-first-garden-that-blooms-with-your-real-pulse-24hi</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/presage-2026-11-04"&gt;Presage SmartSpectra SDK Challenge&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Aura is a small generative garden that wakes up only when Presage SmartSpectra delivers a real human signal.&lt;/strong&gt; A pulse waveform changes the flowers’ glow, sway and bloom; pulse relative to the session baseline shifts the palette. When the signal stops, the flowers freeze rather than performing a fake heartbeat.&lt;/p&gt;

&lt;p&gt;I wanted to make sensing feel less like surveillance and less like another score to optimize. Aura is a one-minute creative wellness break: no streak, diagnosis, account, feed, or microphone. Just a patch of light that responds to you and disappears when you ask it to.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz3tud706r40gm7ckacb4.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz3tud706r40gm7ckacb4.png" alt="Aura rendering real SmartSpectra metrics" width="800" height="678"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/jeffreyturov-dev/aura-garden/blob/main/demo/aura-demo.mp4" rel="noopener noreferrer"&gt;Watch the 69-second demo video (MP4)&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The video shows the consent gate, the recorded-camera replay, the first delayed pulse estimates, changing confidence values, the live relative waveform and the reactive garden. This is not a prerecorded metrics animation: the run observed &lt;strong&gt;251 SmartSpectra metric packets&lt;/strong&gt;, ending with an accumulated &lt;strong&gt;80.1 BPM at 53% confidence&lt;/strong&gt;. The &lt;a href="https://github.com/jeffreyturov-dev/aura-garden/blob/main/demo/run.json" rel="noopener noreferrer"&gt;actual two-second DOM sample log is public&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;For reproducibility without a webcam, the app includes a transformed &lt;strong&gt;synthetic rPPG avatar research clip from UCLA&lt;/strong&gt;, with attribution. It is not footage of a user. The same app also supports a live webcam through custom RGBA frames. Recorded-file mode demonstrates the real SDK pipeline, but I do not present it as proof of accuracy on a live person.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Repository:&lt;/strong&gt; &lt;a href="https://github.com/jeffreyturov-dev/aura-garden" rel="noopener noreferrer"&gt;github.com/jeffreyturov-dev/aura-garden&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The README includes installation, a free Presage API-key setup, unit tests, real-SDK measurement commands, Playwright demo reproduction, and an optional root-free CPU Vulkan setup for headless Linux.&lt;/p&gt;

&lt;h2&gt;
  
  
  How It Responds to Human Signals
&lt;/h2&gt;

&lt;p&gt;Aura uses &lt;code&gt;@smartspectra/node-sdk&lt;/code&gt; &lt;strong&gt;3.4.0&lt;/strong&gt; in two genuine SDK input modes:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// Live browser camera → local WebSocket → SDK custom input&lt;/span&gt;
&lt;span class="nx"&gt;sdk&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;useCustomInput&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;FrameTransform&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;kNone&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="nx"&gt;sdk&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;sendFrame&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;rgba&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;width&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;height&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;width&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;4&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;PixelFormat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;kRGBA&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;timestampUs&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="c1"&gt;// Reproducible recorded camera clip&lt;/span&gt;
&lt;span class="nx"&gt;sdk&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;useFile&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;demoPath&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="na"&gt;interframeDelayMs&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;33&lt;/span&gt; &lt;span class="p"&gt;});&lt;/span&gt;
&lt;span class="nx"&gt;sdk&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;start&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The server requests breathing and cardio bundles and decodes each real metrics buffer:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="nx"&gt;sdk&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;on&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;metrics&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;buffer&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;timestampUs&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;metrics&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;decodeMetrics&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;buffer&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;state&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;metricsState&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;ingest&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;metrics&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="nf"&gt;broadcast&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt; &lt;span class="na"&gt;type&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;metrics&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;timestampUs&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;...&lt;/span&gt;&lt;span class="nx"&gt;state&lt;/span&gt; &lt;span class="p"&gt;});&lt;/span&gt;
&lt;span class="p"&gt;});&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The visual mapping uses:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;cardio.pulseRate&lt;/code&gt;: BPM, confidence and stability display, plus gradual palette/bloom changes relative to a median session baseline;&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;cardio.arterialPressureTrace&lt;/code&gt;: a &lt;strong&gt;relative waveform only&lt;/strong&gt; and the flowers’ instantaneous energy/sway;&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;breathing.upperTrace&lt;/code&gt; and &lt;code&gt;breathing.rate&lt;/code&gt;: additional movement only when a real value exists.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;One detail mattered: SDK windows do not all contain every metric. &lt;code&gt;null&lt;/code&gt; must remain &lt;code&gt;null&lt;/code&gt;. Aura retains the last real sample per signal but never invents a missing pulse or breathing value. If packets become stale, the rendering freezes.&lt;/p&gt;

&lt;h3&gt;
  
  
  The practical issue I hit
&lt;/h3&gt;

&lt;p&gt;The official docs made custom input and file mode look straightforward, but my first cardio run stopped after two seconds with:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;kSetupError: Vulkan is not available on this system.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The fix was not a mock adapter. I ran the SDK with Mesa lavapipe, extracted from pinned Debian packages into a user-writable directory with SHA-256 checks. &lt;code&gt;npm run setup:vulkan&lt;/code&gt; and &lt;code&gt;npm run start:lavapipe&lt;/code&gt; reproduce that root-free setup on Linux x86_64. Native Vulkan is still preferred where available.&lt;/p&gt;

&lt;p&gt;The second timing detail is important for judges: the first pulse can take &lt;strong&gt;12–15 seconds&lt;/strong&gt;. Aura says so in the UI instead of filling the wait with random numbers.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Most Creative or Delightful Use
&lt;/h3&gt;

&lt;p&gt;Aura turns an invisible physiological trace into a small living place rather than a chart. The garden’s visual “breath” comes from actual SDK packets; its restraint is part of the design. There is no simulated garden performing behind the interface while the camera is off.&lt;/p&gt;

&lt;h3&gt;
  
  
  Best Thoughtful Real-World Use
&lt;/h3&gt;

&lt;p&gt;The intended use is a brief, opt-in creative pause for adults who want an ambient experience without creating an account or storing a health history. The controls use plain language, explain the initial delay, expose confidence, and provide one obvious &lt;strong&gt;End session &amp;amp; erase&lt;/strong&gt; action. It deliberately avoids stress labels, diagnoses and medical decisions.&lt;/p&gt;

&lt;h2&gt;
  
  
  Consent and Data Handling
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Consent:&lt;/strong&gt; neither source can start until the user reads the disclosure and ticks an explicit consent checkbox. Starting a source is a second, separate action.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Camera:&lt;/strong&gt; live mode requests &lt;code&gt;getUserMedia({ video, audio: false })&lt;/code&gt;. There is no microphone request, background capture or covert monitoring.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Processing:&lt;/strong&gt; SmartSpectra performs metric computation on-device. Aura sends live RGBA frames only from the browser to its local Node process via the local WebSocket.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Network:&lt;/strong&gt; the SDK requires network access for API-key authentication, subscription validation and encrypted model artifact delivery. Aura does not upload raw camera frames to Presage or any other third party.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Telemetry:&lt;/strong&gt; the SDK is initialized with &lt;code&gt;enableTelemetry: false&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Storage and retention:&lt;/strong&gt; the application has no account, cookie, analytics, database or cloud app backend. Live metrics remain in browser/server memory. Stop releases all media tracks and clears metrics, waveform and garden state.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sharing and deletion:&lt;/strong&gt; there is no automatic export or sharing. &lt;strong&gt;End session &amp;amp; erase&lt;/strong&gt; clears the session. Closing/reloading also releases local media.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Developer recording:&lt;/strong&gt; the separate Playwright script intentionally saves the public demo and its run log. It is never triggered by a normal user session.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Optional AI services:&lt;/strong&gt; none are used at runtime.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Safety:&lt;/strong&gt; Aura is a creative/general-wellness experience for adults. It makes no diagnosis or treatment claim and does not support medical decision-making. The waveform is explicitly labelled relative and never as a blood-pressure reading.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Validation
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;npm test&lt;/code&gt; checks real SDK payload handling, null preservation, erasure, bounded visual mappings and consent-gated controls. The Playwright run is an end-to-end test against the actual SDK and fails unless it sees more than 100 metric packets and a numeric pulse. There were no browser console errors in the checked-in final run.&lt;/p&gt;

&lt;p&gt;Thanks to Presage for the SDK and to UCLA’s Vision Lab for publishing the synthetic rPPG avatar research demonstration used for reproducible testing.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>presagechallenge</category>
      <category>smartspectra</category>
      <category>creativecoding</category>
    </item>
    <item>
      <title>I made 19 Apify Actor examples readable by humans and AI agents</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Mon, 02 Nov 2026 18:13:31 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/i-made-19-apify-actor-examples-readable-by-humans-and-ai-agents-17hk</link>
      <guid>https://dev.to/jeffreyturov/i-made-19-apify-actor-examples-readable-by-humans-and-ai-agents-17hk</guid>
      <description>&lt;p&gt;A useful API is not enough if people and agents cannot evaluate it safely before running it.&lt;/p&gt;

&lt;p&gt;That was the problem I worked on across my 15 public Apify Actors. The tools already had input schemas, README files and pay-per-event pricing. But evaluation still required too much interpretation: which input should I use, what will the output look like, and is this example a real measurement or a technical test?&lt;/p&gt;

&lt;p&gt;I now have 19 public Actor task examples. Each one has two surfaces:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;an HTML page for humans;&lt;/li&gt;
&lt;li&gt;a Markdown page that agents can read without parsing the full application.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;You can browse them from the toolbox:&lt;br&gt;
&lt;a href="https://adagasolutions.lu/toolbox/#examples" rel="noopener noreferrer"&gt;https://adagasolutions.lu/toolbox/#examples&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What a useful public example contains
&lt;/h2&gt;

&lt;p&gt;A good example is not just a saved run with a catchy title. I use this checklist:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;A specific user intent, such as monitoring website AI-readiness changes.&lt;/li&gt;
&lt;li&gt;A small, valid input built from the Actor's current input schema.&lt;/li&gt;
&lt;li&gt;A successful run with a non-empty dataset.&lt;/li&gt;
&lt;li&gt;A dataset view containing the fields the reader actually needs.&lt;/li&gt;
&lt;li&gt;A title and description that match the output.&lt;/li&gt;
&lt;li&gt;A clear dry-run label when the data is only a technical sample.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The last point matters for my AI Brand Visibility Tracker. Its public example does not pretend to measure a real brand. It proves the workflow without calling an AI provider or charging measurement events.&lt;/p&gt;

&lt;p&gt;Here is the example:&lt;br&gt;
&lt;a href="https://apify.com/travelmonitorlab/ai-brand-visibility-tracker/examples/preview-ai-brand-visibility-measurement" rel="noopener noreferrer"&gt;https://apify.com/travelmonitorlab/ai-brand-visibility-tracker/examples/preview-ai-brand-visibility-measurement&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;And its Markdown version:&lt;br&gt;
&lt;a href="https://apify.com/travelmonitorlab/ai-brand-visibility-tracker/examples/preview-ai-brand-visibility-measurement.md" rel="noopener noreferrer"&gt;https://apify.com/travelmonitorlab/ai-brand-visibility-tracker/examples/preview-ai-brand-visibility-measurement.md&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Markdown changes evaluation
&lt;/h2&gt;

&lt;p&gt;An agent can read the example, inspect the input and output, then decide whether the Actor fits the requested workflow. This is more useful than asking it to infer everything from a marketing description.&lt;/p&gt;

&lt;p&gt;It also makes the limitations visible. For example, a brand-visibility result belongs to one model, one prompt set and one run date. It is evidence to compare over time, not a universal ranking.&lt;/p&gt;

&lt;h2&gt;
  
  
  Reliability is still a separate metric
&lt;/h2&gt;

&lt;p&gt;At the time of writing, the public 30-day counters across the 15 Actors show 351 successful runs out of 353. These counters include my own tests and Apify QA. They are useful for reliability monitoring, but they are not customer counts and not proof of recurring adoption.&lt;/p&gt;

&lt;p&gt;I track public examples, run outcomes, Store rankings and reviews separately. Combining them into one vanity metric would hide the real bottleneck.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I am testing next
&lt;/h2&gt;

&lt;p&gt;The technical surface is now complete: 15 public Actors, 19 public examples and a 15-tool MCP facade. The next question is simpler and harder: do these examples help a new user complete a first useful workflow without assistance?&lt;/p&gt;

&lt;p&gt;If you try one, send me the point where you hesitate. That is the part I want to fix.&lt;/p&gt;

</description>
      <category>apify</category>
      <category>mcp</category>
      <category>webscraping</category>
      <category>ai</category>
    </item>
    <item>
      <title>Memoria: Keeping Family Voices Alive with Local Whisper + Gemma</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Tue, 27 Oct 2026 08:17:49 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/memoria-keeping-family-voices-alive-with-local-whisper-gemma-eap</link>
      <guid>https://dev.to/jeffreyturov/memoria-keeping-family-voices-alive-with-local-whisper-gemma-eap</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week4-2026-10-26"&gt;Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Memoria&lt;/strong&gt; is a local oral-history archivist. You give it a family voice memo. It returns a structured keepsake containing:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the complete transcript with timestamps;&lt;/li&gt;
&lt;li&gt;the people, places, and time references explicitly mentioned;&lt;/li&gt;
&lt;li&gt;themes and a short summary;&lt;/li&gt;
&lt;li&gt;verbatim quotes checked against the transcript;&lt;/li&gt;
&lt;li&gt;three thoughtful questions to ask during the next conversation.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The result is exported as JSON and Markdown. There is also a small web interface with a playable demo recording.&lt;/p&gt;

&lt;p&gt;I built it for the &lt;strong&gt;Ghost in the Machine&lt;/strong&gt; theme as a way to honor the past without surrendering it to a cloud service. A recording of a grandparent is not ordinary prompt data. It can contain names, addresses, health details, and stories that a family never intended to publish.&lt;/p&gt;

&lt;p&gt;Memoria runs both speech recognition and language understanding locally. The recording and transcript never leave the machine for inference.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://sometimes-phantom-academy-adequate.trycloudflare.com" rel="noopener noreferrer"&gt;Open the live Memoria demo&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Press &lt;strong&gt;“Archive this memory”&lt;/strong&gt; to run the complete pipeline on local CPU models. Please allow about a minute.&lt;/p&gt;

&lt;p&gt;The bundled 58-second story is &lt;strong&gt;synthetic and fictional&lt;/strong&gt;, narrated with TTS specifically so that no private family recording is exposed. The processing is not mocked: the button runs Whisper and Gemma against that audio. The demo tunnel is temporary; the repository contains everything needed to run it locally.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/jeffreyturov-dev/memoria-local" rel="noopener noreferrer"&gt;github.com/jeffreyturov-dev/memoria-local&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The repository includes the CLI, FastAPI demo, synthetic demo audio, generated example output, tests, and setup instructions.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;The pipeline is deliberately small and inspectable:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;voice memo
    │
    ▼
faster-whisper / Whisper small (CPU int8)
    │  transcript + segment timestamps
    ▼
Gemma 3 4B IT QAT through llama.cpp
    │  strict JSON memory card
    ▼
deterministic quote validation
    │
    ├── JSON archive
    ├── Markdown keepsake
    └── local web result
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Whisper handles the voice. Gemma receives only the resulting transcript and must produce eight fields: title, summary, people, places, time references, themes, quotes, and follow-up questions. Temperature is zero, missing evidence becomes an empty array, and the prompt explicitly forbids adding facts.&lt;/p&gt;

&lt;p&gt;A small model can still produce a quotation that sounds plausible but does not occur in the recording. For an archive, that is unacceptable. I therefore added a deterministic guard after generation:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;verify_quotes&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;card&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;transcript&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;normalized_transcript&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;normalize_words&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;transcript&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;verified&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[]&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;quote&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;card&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;quotes&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]:&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="nf"&gt;isinstance&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;quote&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;str&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="nf"&gt;normalize_words&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;quote&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;normalized_transcript&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="n"&gt;verified&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;quote&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;strip&lt;/span&gt;&lt;span class="p"&gt;())&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;verified&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Only whitespace and letter case are normalized. If the passage is absent, Memoria drops it rather than presenting it as the speaker's words. The original transcript always remains beside the generated card, so a family can review the source rather than trusting the summary blindly.&lt;/p&gt;

&lt;p&gt;The demo uses FastAPI, a concurrency lock, and a 25 MB limit for local audio uploads. No API key or account is required.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Moment It Landed
&lt;/h2&gt;

&lt;p&gt;The fictional demo story describes a grandfather's radio repair shop, yellow marigolds, cinnamon tea, and a rebuilt red bicycle called &lt;em&gt;La Valiente&lt;/em&gt;. Memoria recovered those details and preserved this passage verbatim:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;A broken thing is only a story waiting for patient hands.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;It then suggested questions such as asking what the radios sounded like and why the bicycle had that name. That was the moment the project stopped feeling like a transcription wrapper: the useful output is not just what was said, but a gentle prompt to record the next chapter while you still can.&lt;/p&gt;

&lt;p&gt;The recording is synthetic, and I have kept that distinction visible in the demo and repository. A real family archive should start with consent and human review, not an invented claim of authenticity.&lt;/p&gt;

&lt;h2&gt;
  
  
  Verified Results
&lt;/h2&gt;

&lt;p&gt;I ran the project rather than stopping at a screenshot or a plausible demo response:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Check&lt;/th&gt;
&lt;th&gt;Actual result&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Unit tests&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;6/6 passed&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Demo audio duration&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;57.79 seconds&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Full CLI run, audio → archive&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;33.36 seconds&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Public web run through the tunnel&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;29.47 seconds&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Whisper language detection&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;English, 0.9993 probability&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Displayed quotes&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;3/3 found in the transcript&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;For the offline-mode run, I forced &lt;code&gt;HTTP_PROXY&lt;/code&gt;, &lt;code&gt;HTTPS_PROXY&lt;/code&gt;, and &lt;code&gt;ALL_PROXY&lt;/code&gt; to a dead localhost address, with only localhost excluded. The complete cached-model pipeline still produced the transcript and memory card. The generated example is committed under &lt;code&gt;demo/output/&lt;/code&gt; for inspection.&lt;/p&gt;

&lt;p&gt;The unit tests cover fenced JSON parsing, missing-field rejection, hallucinated-quote rejection, whitespace normalization, human-readable people fields, and preservation of the original transcript.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Open Innovation Matters
&lt;/h2&gt;

&lt;p&gt;Open weights and open-source inference are not decorative details here; they make the product possible.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Privacy:&lt;/strong&gt; a closed transcription or LLM API would send intimate family material to infrastructure the family does not control. Memoria performs both stages locally.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Longevity:&lt;/strong&gt; an oral-history archive should not become inaccessible because a vendor retired a model endpoint. The code is MIT-licensed, the archive is plain JSON and Markdown, and the inference components can be replaced.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Offline use:&lt;/strong&gt; after the models are downloaded, the pipeline works without cloud AI access. That makes it usable in the places where families actually sit down together, not only where connectivity and a subscription are available.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Zero marginal inference bill:&lt;/strong&gt; there is no per-minute or per-token charge for a longer conversation. The trade-off is CPU latency and the need to review a small model's work.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;p&gt;Memoria does not independently fact-check a summary, and a high language-detection score does not guarantee a perfect transcript. Names can be misheard; in the demo, the source script says “Mateo” while Whisper transcribes “Matteo,” and the card faithfully follows that transcript. This is exactly why the source text and human review remain essential. Only record and archive someone else's story with their consent.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; Gemma 3 4B IT QAT performs the local structured memory extraction through llama.cpp.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Overall:&lt;/strong&gt; a complete, tested, locally runnable oral-history tool with a real public demo.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Slow and Steady: An Agent That Runs My Apify Side Business While I Sleep</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Sat, 24 Oct 2026 08:12:57 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/slow-and-steady-an-agent-that-runs-my-apify-side-business-while-i-sleep-2j5p</link>
      <guid>https://dev.to/jeffreyturov/slow-and-steady-an-agent-that-runs-my-apify-side-business-while-i-sleep-2j5p</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;I run 13 public web-scraping Actors on the Apify Store as a side business. Keeping up with the platform — blog announcements, Store trends, pricing-compliance deadlines, new monetization features — is exactly the kind of chore I kept putting off. So I built &lt;strong&gt;Phoenix&lt;/strong&gt;: an autonomous agent on a cron schedule that does the whole daily platform watch for me and delivers an 8-line report while I'm still asleep.&lt;/p&gt;

&lt;p&gt;Every morning, on its own, it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;scans the Apify blog RSS for anything published in the last 48 hours&lt;/li&gt;
&lt;li&gt;pulls the Apify Store top-10 to spot demand trends&lt;/li&gt;
&lt;li&gt;audits my Actors' pricing configuration (pay-per-event migration compliance)&lt;/li&gt;
&lt;li&gt;checks my saved tasks against the Store's newest SEO surface, published task examples&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;executes whatever is deliverable end-to-end&lt;/strong&gt; — not just reading, but doing: configs, fixes, submissions — then logs done-vs-read in a dated markdown file&lt;/li&gt;
&lt;li&gt;watches dev.to challenges with cash prizes and drafts submissions (yes, the research for this very post was done that way)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The sloth philosophy, applied: conserve energy, let the machine do the rounds.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;Real, unedited output from this morning's run (October 24, 2026):&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Blog scan:&lt;/strong&gt; Apify shipped bulk task-example publishing via API on Oct 22 — their case study claims organic traffic to an Actor page doubled after publishing examples.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Audit:&lt;/strong&gt; I have 11 saved tasks across my Actors and &lt;strong&gt;zero published task examples&lt;/strong&gt;. Gap identified, publish payloads prepared.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Action attempted:&lt;/strong&gt; the agent tried the publish API call… and got &lt;code&gt;403 platform-feature-disabled&lt;/code&gt;. Instead of silently failing, it traced the cause: my account had burned through its &lt;strong&gt;$5 monthly hard limit&lt;/strong&gt; — 0.58 GB of residential proxy (~$4.60) consumed by one overeager scraper run. It flagged it for my review, with the numbers.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I learned about a new Store SEO feature &lt;em&gt;and&lt;/em&gt; a billing problem before my first coffee, with zero clicks.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;The harness is &lt;a href="https://github.com/NousResearch/hermes-agent" rel="noopener noreferrer"&gt;Hermes Agent&lt;/a&gt; — an open-source agent runtime by Nous Research (&lt;a href="https://hermes-agent.nousresearch.com/docs" rel="noopener noreferrer"&gt;docs&lt;/a&gt;). The daily watch is a cron job that hands the agent a task file; the agent loops: scan → reason → act → log. The platform knowledge lives in plain-markdown "skill" files, and the actual checks are small scripts the agent writes and runs itself, like this abridged one from today:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;urllib&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;getj&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;url&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;req&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;urllib&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Request&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;url&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;headers&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;User-Agent&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;phoenix-cron&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;})&lt;/span&gt;
    &lt;span class="k"&gt;with&lt;/span&gt; &lt;span class="n"&gt;urllib&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;urlopen&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;req&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;timeout&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;30&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;load&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="c1"&gt;# Which of my Actors are still PPE-compliant?
&lt;/span&gt;&lt;span class="n"&gt;acts&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;getj&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://api.apify.com/v2/acts?my=1&amp;amp;token=&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;TOKEN&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt;&amp;amp;limit=50&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;it&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;acts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;data&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;items&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]:&lt;/span&gt;
    &lt;span class="n"&gt;det&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;getj&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://api.apify.com/v2/acts/&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;it&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;id&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt;?token=&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;TOKEN&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;data&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;det&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;isPublic&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="nf"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;det&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;name&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;-&amp;gt;&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;det&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;pricingInfos&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="ow"&gt;or&lt;/span&gt; &lt;span class="p"&gt;[{}])[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;pricingModel&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No framework lock-in: the harness calls shell, Python, and a real browser over CDP. That's the whole stack.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Orchestration:&lt;/strong&gt; Hermes Agent, open-source. The agent reads its instructions, picks tools, writes its own helper scripts, and iterates until the task verifies — a real agent loop, not a Zapier-style flow.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Scheduling:&lt;/strong&gt; a boring, glorious cron line. Slow and steady, every day at the same hour.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Knowledge:&lt;/strong&gt; markdown skill files (actor doctrine, preflight checks, the publishing API contract). When Apify changed task-example limits from 50 to 10 per Actor, I edited a text file — not a support ticket.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model-agnostic:&lt;/strong&gt; the harness doesn't care which LLM drives it. You can point the same setup at an open-weight model and the loop still works; the automation survives the swap.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Three concrete things a closed automation platform wouldn't give me:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;My credentials never leave my machine.&lt;/strong&gt; The agent touches Gmail, Apify, and Stripe tokens daily. They live in local files on my own server in Luxembourg — not in a vendor's cloud, not in someone's "workspace".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Everything is inspectable and forkable.&lt;/strong&gt; Skills are text. Scripts are text. Logs are text. When something breaks at 6 AM, I can diff it, fix it, and the fix is permanent — not a black-box retry policy I can't see.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No platform risk on the orchestration layer.&lt;/strong&gt; Closed assistants lock you into their model, their pricing, their roadmap. An open harness means my morning routine keeps working even if I switch models tomorrow. Open weights crossing the quality bar is a when, not an if — and I'm already wired for it.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The Lazy Payoff
&lt;/h2&gt;

&lt;p&gt;About &lt;strong&gt;45 minutes of daily platform-watch compressed into reading 8 lines&lt;/strong&gt;. This week alone the agent prepared task-example configs for the Store's new SEO surface, verified pay-per-event pricing on all 13 public Actors ahead of the migration deadline, and caught a hard-limit billing issue — all while I did literally nothing.&lt;/p&gt;

&lt;p&gt;Sloths had it figured out. Build the thing once, let it do the rounds forever.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Sloth Digest 🦥 — a local Gemma 3 agent that does my inbox triage so I don't have to</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Tue, 20 Oct 2026 08:10:45 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/sloth-digest-a-local-gemma-3-agent-that-does-my-inbox-triage-so-i-dont-have-to-9gk</link>
      <guid>https://dev.to/jeffreyturov/sloth-digest-a-local-gemma-3-agent-that-does-my-inbox-triage-so-i-dont-have-to-9gk</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Sloth Digest&lt;/strong&gt; is a local AI agent that reads my unread Gmail overnight and hands me a morning digest, so I never have to open a cold inbox again.&lt;/p&gt;

&lt;p&gt;Every email gets triaged by an &lt;strong&gt;open-weight model running entirely on my own machine&lt;/strong&gt; (Gemma 3 4B IT, QAT-quantized, served by llama.cpp on plain CPU) into three buckets:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;⚡ &lt;strong&gt;Needs you&lt;/strong&gt; — reply, pay, confirm, decide — each with a suggested one-line action&lt;/li&gt;
&lt;li&gt;👀 &lt;strong&gt;Worth 10 seconds&lt;/strong&gt; — FYI, no action needed&lt;/li&gt;
&lt;li&gt;😴 &lt;strong&gt;Skip safely&lt;/strong&gt; — promos, newsletters, notification noise&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Who gets to be lazier? Me, every single morning. Instead of 15–30 minutes of scanning subject lines and context-switching, I read one short markdown page and only touch the 3–4 emails that actually need a human.&lt;/p&gt;

&lt;p&gt;Here's real output from the bundled sample inbox:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="gh"&gt;# 🦥 Sloth Digest — morning triage&lt;/span&gt;

&lt;span class="gu"&gt;## ⚡ Needs you (4)&lt;/span&gt;
&lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="gs"&gt;**Energia Lux**&lt;/span&gt; — Your October invoice — €84.12 due Oct 27
  Electricity bill of €84.12 is due soon.
  → Pay the invoice before October 27th.
&lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="gs"&gt;**Marco Silva**&lt;/span&gt; — Re: Thursday moved to 15:00?
  Plumber availability needs confirmation for Thursday.
  → Confirm 15:00 or propose an alternative time.
...

&lt;span class="gu"&gt;## 😴 Skip safely (1)&lt;/span&gt;
&lt;span class="p"&gt;-&lt;/span&gt; &lt;span class="gs"&gt;**ShopMega**&lt;/span&gt; — 🔥 70% OFF ends tonight — don't miss out!
  Promotional email about a limited-time sale.

6 emails triaged locally in 14s (~12 min of manual triage saved).
Nothing left this machine.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Live demo:&lt;/strong&gt; &lt;a href="https://yoga-unexpected-taylor-provide.trycloudflare.com" rel="noopener noreferrer"&gt;https://yoga-unexpected-taylor-provide.trycloudflare.com&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It shows the sample inbox (6 fake emails — bills, a friend asking for a confirmation, a flash sale, a delivery notice, a medical appointment reminder) and runs the triage live on the local Gemma model when you hit the button. ~14 seconds on CPU, no cloud API involved anywhere.&lt;/p&gt;

&lt;p&gt;You can reproduce the whole thing offline:&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;./llama-server &lt;span class="nt"&gt;-m&lt;/span&gt; gemma-3-4b-it-qat-Q4_0.gguf &lt;span class="nt"&gt;--port&lt;/span&gt; 8801
python sloth_digest.py &lt;span class="nt"&gt;--demo&lt;/span&gt;   &lt;span class="c"&gt;# sample inbox, zero credentials needed&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/jeffreyturov-dev" rel="noopener noreferrer"&gt;
        jeffreyturov-dev
      &lt;/a&gt; / &lt;a href="https://github.com/jeffreyturov-dev/sloth-digest" rel="noopener noreferrer"&gt;
        sloth-digest
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Local AI agent that triages your Gmail overnight into a morning digest — Gemma 3 4B, 100% offline, privacy-first
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;🦥 Sloth Digest&lt;/h1&gt;
&lt;/div&gt;

&lt;p&gt;A local AI agent that reads your unread Gmail overnight and hands you a
morning digest — so you don't have to.&lt;/p&gt;
&lt;p&gt;Every email is triaged into three buckets by an &lt;strong&gt;open-weight model running
entirely on your own machine&lt;/strong&gt; (Gemma 3 4B IT QAT via llama.cpp):&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;⚡ &lt;strong&gt;Needs you&lt;/strong&gt; — reply, pay, confirm, decide (with a suggested action)&lt;/li&gt;
&lt;li&gt;👀 &lt;strong&gt;Worth 10 seconds&lt;/strong&gt; — FYI, no action&lt;/li&gt;
&lt;li&gt;😴 &lt;strong&gt;Skip safely&lt;/strong&gt; — promos, newsletters, notification noise&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;No email content ever leaves your computer. Works offline. Costs nothing
per run — set it up once and forget it.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Why local?&lt;/h2&gt;
&lt;/div&gt;
&lt;p&gt;Your inbox is the most private dataset you own. A closed API means sending
every message — invoices, medical appointments, personal threads — to
someone else's servers. Sloth Digest runs the whole pipeline against a
local OpenAI-compatible endpoint, so the "open" piece &lt;em&gt;is&lt;/em&gt; the product:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy&lt;/strong&gt;: nothing leaves the…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/jeffreyturov-dev/sloth-digest" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The agent is ~200 lines of Python stdlib. The only dependencies are llama.cpp (for inference) and, for the optional web UI, FastAPI.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;The architecture is deliberately boring — a sloth would approve:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;unread Gmail ──► fetch (Gmail API, strictly read-only)
                   │
                   ▼
        ┌─────────────────────┐
        │ Gemma 3 4B IT (QAT) │  per email: CATEGORY / SUMMARY / ACTION
        │ llama.cpp, CPU      │  strict 3-line verdict, temperature 0
        └─────────────────────┘
                   │
                   ▼
        morning digest (markdown, grouped by urgency)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Three design choices make a small 4B model reliable enough to trust with triage:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Rigid output contract.&lt;/strong&gt; The model must answer with exactly three lines (&lt;code&gt;CATEGORY&lt;/code&gt; / &lt;code&gt;SUMMARY&lt;/code&gt; / &lt;code&gt;ACTION&lt;/code&gt;) at temperature 0. No essays, no JSON to break.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Graceful degradation.&lt;/strong&gt; Anything unparseable lands in "check manually" instead of being silently misfiled. A triage tool that hides emails is worse than no tool.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Read-only by construction.&lt;/strong&gt; The agent fetches unread mail and prints a digest. It never marks, archives, or deletes anything — the worst it can do is be wrong on a page you skim anyway.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Local inference is not a gimmick here, it's the point: the whole pipeline — fetch, classify, summarize — runs against llama.cpp's OpenAI-compatible endpoint on &lt;code&gt;127.0.0.1&lt;/code&gt;. Pull the network cable and it still works.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Your inbox is probably the most private dataset you own: invoices, medical appointments, family threads, business negotiations. Triage means reading &lt;em&gt;all of it&lt;/em&gt; — and with a closed API, "reading" means shipping every message to someone else's servers, on someone else's terms, under a privacy policy that can change.&lt;/p&gt;

&lt;p&gt;Because Gemma's weights are open and llama.cpp is open source, the equation flips:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy by architecture, not by promise.&lt;/strong&gt; Nothing leaves the machine, so there's nothing to leak, log, or train on.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Works offline.&lt;/strong&gt; Plane, cabin, outage — the digest still runs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Free forever.&lt;/strong&gt; No per-token bill. I set it up once and it costs me exactly zero per run, forever. That's the sloth economics a closed API can never offer.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The Lazy Payoff
&lt;/h2&gt;

&lt;p&gt;I ran it on my real unread inbox this morning: &lt;strong&gt;8 emails triaged in ~20 seconds&lt;/strong&gt;, and the digest correctly separated two things I genuinely needed to act on from the newsletter noise. Manual triage averages ~2 minutes per email once you include the context-switching — so the agent saves me roughly &lt;strong&gt;15 minutes a day, ~90 hours a year&lt;/strong&gt;, for a one-time setup cost and €0 in API fees.&lt;/p&gt;

&lt;p&gt;The cron job runs at 7am. I wake up, read one page, and go back to being lazy. The sloth wins. 🦥&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt; — Gemma 3 4B IT (QAT) is the entire brain of the agent, running locally via llama.cpp.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>I Let an AI Agent Run Parts of My Real Businesses for a Year. Here Are the Receipts.</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Mon, 19 Oct 2026 21:02:13 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/i-let-an-ai-agent-run-parts-of-my-real-businesses-for-a-year-here-are-the-receipts-4lkn</link>
      <guid>https://dev.to/jeffreyturov/i-let-an-ai-agent-run-parts-of-my-real-businesses-for-a-year-here-are-the-receipts-4lkn</guid>
      <description>&lt;p&gt;One year ago, I opened a terminal on a small home server and gave an AI agent its first real task.&lt;/p&gt;

&lt;p&gt;I'm not a developer. I run a small hotel in Mersch, Luxembourg. I had no team, no budget for an agency, and a long list of things that needed to exist: a booking site, a phone receptionist, payments, a second company, SEO, ads.&lt;/p&gt;

&lt;p&gt;Today, that agent — the same one, still living in the same terminal — runs a real slice of my entrepreneur life. Not demos. Not chat screenshots. Production systems, with real customers and real money.&lt;/p&gt;

&lt;p&gt;This is episode 1 of an 8-part series. Everything I claim here is publicly verifiable, and I've linked the receipts.&lt;/p&gt;

&lt;h2&gt;
  
  
  What "running a business with an AI agent" actually means
&lt;/h2&gt;

&lt;p&gt;Forget the viral stories about AIs asking for donations. Here's the boring, real version:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It answers my hotel's phone.&lt;/strong&gt; A voice AI receptionist picks up real calls, checks availability, quotes our direct rate (10% below the OTAs), and takes bookings. Guests talk to it daily. It announces call recording, because compliance isn't optional.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It filed my second company.&lt;/strong&gt; Statutes, government forms, notary coordination, registry filing. The ministerial authorization for my new company (a SARL-S, the Luxembourg equivalent of an LLC) arrived this month. The AI drafted; I signed. That division of labor matters, and I'll come back to it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It published an MCP server on the official registry.&lt;/strong&gt; Any AI agent in the world can now discover my hotel federation and book real rooms with 0% OTA commission — versus the 15-25% the platforms take. It's live on Smithery and Glama too.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It runs the money plumbing.&lt;/strong&gt; Stripe Connect charges, platform fees, payouts, a revenue dashboard with real processing fees computed per merchant. Manual payouts — because I keep a human hand on money movements. On purpose.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It watches itself.&lt;/strong&gt; Dozens of cron jobs, watchdogs, and health checks. When Booking.com started captcha-ing my server's IP mid-October, the monitoring caught it and the agent rebuilt the pipeline through a different path within a day.&lt;/p&gt;

&lt;h2&gt;
  
  
  The receipts
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;My hotel's live booking site: &lt;a href="https://yashahotel.com" rel="noopener noreferrer"&gt;yashahotel.com&lt;/a&gt; (Stripe Connect in production)&lt;/li&gt;
&lt;li&gt;My company's site: &lt;a href="https://adagasolutions.lu" rel="noopener noreferrer"&gt;adagasolutions.lu&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The MCP server: &lt;a href="https://smithery.ai/servers/@jeffreyturov/adaga-network" rel="noopener noreferrer"&gt;smithery.ai/servers/@jeffreyturov/adaga-network&lt;/a&gt; — also on the official MCP registry&lt;/li&gt;
&lt;li&gt;The technical write-up of the MCP build: &lt;a href="https://dev.to/jeffreyturov/i-built-an-mcp-server-that-lets-ai-agents-book-real-hotel-rooms-no-ota-commission-2g54"&gt;my previous article here on dev.to&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;My hotel company exists in the Luxembourg trade register since 2005 — the AI part is what happened in the last 12 months&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What the agent does alone vs. what I validate
&lt;/h2&gt;

&lt;p&gt;This is the part most "AI entrepreneur" stories skip, so let me be precise:&lt;/p&gt;

&lt;p&gt;The agent does, end-to-end and without asking: research, code, deployment to my sandbox, testing, monitoring, documentation, drafting emails, answering comments on my articles, scraping, data pipelines.&lt;/p&gt;

&lt;p&gt;I validate, always: external communications to institutions, anything touching money movements, client-facing promises, legal signatures, and publishing on my personal accounts.&lt;/p&gt;

&lt;p&gt;That boundary isn't a limitation of the AI. It's a governance choice, and it's exactly why this still works after a year: the agent is fast, and I am the brakes.&lt;/p&gt;

&lt;h2&gt;
  
  
  The failures are the best part
&lt;/h2&gt;

&lt;p&gt;I'll dedicate a full episode to this, because it's where the real learning is: the agent once reverted my production site by deploying from an outdated copy. Google revoked my Gmail tokens because of a password reset and killed half my integrations. Booking.com banned my server IP. A cron job nearly got me disqualified from a dev challenge by double-submitting.&lt;/p&gt;

&lt;p&gt;Every single one of those failures is now a documented procedure the agent will never repeat. That — not the model, not the prompts — is the actual moat.&lt;/p&gt;

&lt;h2&gt;
  
  
  The series
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Ep.2&lt;/strong&gt; — The AI receptionist (with audio)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ep.3&lt;/strong&gt; — The AI that filed my company (with the documents)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ep.4&lt;/strong&gt; — Strangers' AI agents booking my rooms (the MCP deep-dive)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ep.5&lt;/strong&gt; — Everything that broke (the viral one, probably)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ep.6&lt;/strong&gt; — The money: honest ranges&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ep.7&lt;/strong&gt; — The machine: 1 human, ~100 learned procedures, 0 employees&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ep.8&lt;/strong&gt; — One-year review + the book&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I'm writing this for two kinds of people: developers who want to see what a year of agentic operations actually looks like under the hood, and small business owners who suspect they can't afford an agency — you're right, and there's another way now.&lt;/p&gt;

&lt;p&gt;Which episode do you want first? Tell me in the comments — I'll adjust the order.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>automation</category>
      <category>agents</category>
      <category>buildinpublic</category>
    </item>
    <item>
      <title>I built an MCP server that lets AI agents book real hotel rooms (no OTA commission)</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Fri, 16 Oct 2026 15:26:46 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/i-built-an-mcp-server-that-lets-ai-agents-book-real-hotel-rooms-no-ota-commission-2g54</link>
      <guid>https://dev.to/jeffreyturov/i-built-an-mcp-server-that-lets-ai-agents-book-real-hotel-rooms-no-ota-commission-2g54</guid>
      <description>&lt;p&gt;Booking.com takes 15-25% commission and owns the customer relationship. I run a small&lt;br&gt;
network of independent establishments in Luxembourg, and I wanted AI agents to be able to&lt;br&gt;
book them directly — machine-native, at the establishment's exact direct price.&lt;/p&gt;

&lt;p&gt;So I built &lt;strong&gt;Adaga Network — Direct Booking Federation&lt;/strong&gt;: one MCP endpoint that fans out&lt;br&gt;
to every establishment in the network.&lt;/p&gt;

&lt;h2&gt;
  
  
  The architecture
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Each establishment runs the same open template: an &lt;code&gt;mcp.js&lt;/code&gt; module mounted on &lt;code&gt;/mcp&lt;/code&gt;
inside its existing Express app. No new port, no new service — a pure proxy to the
app's own REST API.&lt;/li&gt;
&lt;li&gt;The federator (&lt;a href="https://adagasolutions.lu/network/mcp" rel="noopener noreferrer"&gt;adagasolutions.lu/network/mcp&lt;/a&gt;) is a Streamable HTTP stateless MCP
server that keeps a registry of tenants and calls their MCP tools in parallel
(6s timeout per tenant, dead tenants are skipped and reported).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;3 tools cover the whole journey:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tool&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;network_directory&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Lists every establishment (brand, type, city, slug)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;network_search&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Fan-out live availability + exact total prices, sorted&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;network_book&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Creates a REAL booking, hands secure card payment to the human&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  The payment hand-off is a feature
&lt;/h2&gt;

&lt;p&gt;The agent never touches card data. &lt;code&gt;network_book&lt;/code&gt; returns payment instructions; the human&lt;br&gt;
completes a Stripe Checkout; unpaid bookings expire automatically. Booking confirmed only&lt;br&gt;
after payment — and a 409 if the room became unavailable in between.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I learned writing tool descriptions
&lt;/h2&gt;

&lt;p&gt;The tool description is the #1 selection criterion for agents. After running the tools&lt;br&gt;
through a TDQS audit, the fixes that mattered:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;State auth requirements, rate limits and pagination explicitly (mine: none, none, none)&lt;/li&gt;
&lt;li&gt;Put the full workflow in the transactional tool's description&lt;/li&gt;
&lt;li&gt;Describe every optional input format (phone: international format, extras: integer ids)&lt;/li&gt;
&lt;li&gt;Units everywhere: prices in euro cents, dates YYYY-MM-DD, max 30 nights&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Distribution
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Official MCP Registry: &lt;code&gt;io.github.jeffreyturov-dev/adaga-network&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Glama connector with health checks (hourly) — a watchdog pings the endpoint every 10 min&lt;/li&gt;
&lt;li&gt;The honest pitch: no OTA commission — the traveler pays the exact direct price,
the establishment keeps its customer.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Salons and restaurants join the network next. If you run an independent establishment and&lt;br&gt;
want AI agents to book you directly, the onboarding template takes 2-4 hours.&lt;/p&gt;

&lt;h2&gt;
  
  
  Live links
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Network page for establishments: &lt;a href="https://adagasolutions.lu/agents-ia/" rel="noopener noreferrer"&gt;adagasolutions.lu/agents-ia&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;First federated hotel (live direct booking): &lt;a href="https://yashahotel.com/" rel="noopener noreferrer"&gt;yashahotel.com&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The hotel's own agent page: &lt;a href="https://yashahotel.com/agent/" rel="noopener noreferrer"&gt;yashahotel.com/agent&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>mcp</category>
      <category>ai</category>
      <category>node</category>
      <category>travel</category>
    </item>
    <item>
      <title>I built an offline review-reply concierge for my family's hotel — €0 per reply, five languages, no cloud</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Tue, 13 Oct 2026 09:00:08 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/i-built-an-offline-review-reply-concierge-for-my-familys-hotel-eu0-per-reply-five-languages-no-bol</link>
      <guid>https://dev.to/jeffreyturov/i-built-an-offline-review-reply-concierge-for-my-familys-hotel-eu0-per-reply-five-languages-no-bol</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;review-concierge&lt;/strong&gt; drafts thoughtful, multilingual replies to guest reviews — entirely offline, on the machine already sitting at the hotel's front desk, at a marginal cost of exactly €0 per reply.&lt;/p&gt;

&lt;p&gt;Who pays for this? Small independent hotels. My family runs one: a 28-room hotel near Luxembourg City's train station, with 280+ Google reviews written in five languages (French, German, English, Dutch, Portuguese). Replying to reviews is one of the highest-ROI tasks in hospitality — and the closed-source answer to it, &lt;a href="https://www.mara-solutions.com/pricing" rel="noopener noreferrer"&gt;MARA&lt;/a&gt;, costs €70–100 per property per month. That's a real line item for a family business. The core of that job — a well-written, on-brand draft in the guest's own language — is something an open-weight model now does for free.&lt;/p&gt;

&lt;p&gt;The side-hustle angle cuts both ways: it's a tool a small business &lt;em&gt;would&lt;/em&gt; pay for, and it's only viable &lt;em&gt;because&lt;/em&gt; open weights make the marginal cost zero.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;🟢 Live demo (temporary tunnel up for the judging period): &lt;a href="https://register-annie-sites-saved.trycloudflare.com" rel="noopener noreferrer"&gt;https://register-annie-sites-saved.trycloudflare.com&lt;/a&gt;&lt;/strong&gt; — paste a guest review in any language, watch the reply draft locally on plain CPU. Nothing you type leaves the machine it runs on.&lt;/p&gt;

&lt;p&gt;Real run, real review (one of our actual 3★ Google reviews, mentioning a broken shower holder and mould in the bathroom):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;$ python reply.py --file reviews_yasha_sample.csv --out replies_yasha.csv
[3/10] done (438 chars)

→ We're so sorry to hear about the issues you experienced with the bathroom
  and shower head, and thank you for pointing out the broken holder – we
  appreciate you finding a resourceful solution! We sincerely apologize that
  the air-conditioning wasn't working as advertised and for the inconvenience
  this caused. We're immediately addressing the mildew and mould, and
  replacing the shower head holder. Thank you for your feedback; we truly
  value it and hope to welcome you back for a much better experience.

  The Yasha Hotel team
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Notice what it did: acknowledged the &lt;em&gt;specific&lt;/em&gt; problems, thanked the guest instead of arguing, named a concrete fix — and did it in English, because the review was in English. The French reviews get French replies, the German ones German, the Dutch one Dutch. I verified all 10 drafts against 10 real reviews: &lt;strong&gt;10/10 in the correct language, 10/10 specific, zero arguments with guests.&lt;/strong&gt; The full input/output table is in the repo's README.&lt;/p&gt;

&lt;p&gt;There's also a tiny web UI (&lt;code&gt;app.py&lt;/code&gt;, ~70 lines) so a non-technical front-desk person can paste a review and get a draft in ~20 seconds — no terminal required.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Code: &lt;a href="https://codeberg.org/veysel-devia/review-concierge" rel="noopener noreferrer"&gt;codeberg.org/veysel-devia/review-concierge&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The core is &lt;code&gt;reply.py&lt;/code&gt; — ~120 lines, Python standard library only. It talks to a local &lt;a href="https://github.com/ggml-org/llama.cpp" rel="noopener noreferrer"&gt;llama.cpp&lt;/a&gt; server running &lt;strong&gt;Gemma 3 4B (QAT, Q4_0)&lt;/strong&gt;, Google's open-weight model, on a plain CPU (no GPU, ~12 tok/s, a full draft in ~20 s).&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;Two model calls per draft, both against the local OpenAI-compatible endpoint:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Language detection&lt;/strong&gt; — the model itself classifies the review (temperature 0, ISO 639-1 out). No extra dependency, no langid/fasttext baggage.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reply drafting&lt;/strong&gt; — a system prompt that encodes a reputation-manager playbook: reply &lt;em&gt;entirely&lt;/em&gt; in the detected language (hard constraint, restated in the user message), 40–80 words, acknowledge one specific detail, apologise without arguing, name a concrete fix for maintenance issues, never discuss the neighbourhood, sign as the hotel.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The one real bug I hit is instructive: my first version said "reply in the same language as the review" — and Gemma 4B, primed by the Luxembourg hotel context, answered roughly half the English reviews in French. The fix was making the constraint explicit and mechanical (detect first, then inject "the review is in &lt;em&gt;English&lt;/em&gt;, reply in &lt;em&gt;English&lt;/em&gt;"). Small-model prompting is engineering, not vibes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Three reasons, and they're the whole business case:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;€0 marginal cost is the product.&lt;/strong&gt; A review-reply SaaS charges per property per month because every reply costs them an API call. With open weights, the hundredth reply costs exactly as much as the first: nothing. That's what turns "nice AI demo" into "tool a 28-room hotel can actually afford."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Guest data never leaves the building.&lt;/strong&gt; Reviews contain personal data — names, stay dates, occasionally health details in complaints. Under GDPR, piping them through a third-party API is a compliance conversation most small hoteliers don't want to have. Local inference means there is no conversation: the data stays on the machine, verified working with all egress blocked.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The voice is a text file, not a paywall.&lt;/strong&gt; The hotel's tone, apology policy and signature live in a system prompt the owner can edit. With closed tools, that's an enterprise tier.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  My First User
&lt;/h2&gt;

&lt;p&gt;The first user is my family's hotel — which is also how I knew what to build. I ran the tool against 10 of its real, currently-unanswered Google reviews (the exact mix a small property faces: a glowing 5★ from a returning business traveller, a furious 1★ in French about a mattress, a 3★ in Dutch). The drafts it produced are good enough to post after a 30-second human read — which is the correct workflow anyway: the AI drafts, the human owns the reply. Would they pay for it? They already pay with something scarcer than money: the reviews were sitting unanswered because nobody at the front desk has time to craft careful replies in five languages. Now the time cost is a paste and a skim.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt; — the entire product runs on Gemma 3 4B, Google's open-weight model, served locally through llama.cpp. Open weights are not an implementation detail here; they're what makes the economics and the privacy story work.&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>I built a free AEO checker — is your website visible to ChatGPT, Perplexity and Google AI?</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Sat, 10 Oct 2026 01:51:48 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/i-built-a-free-aeo-checker-is-your-website-visible-to-chatgpt-perplexity-and-google-ai-10gl</link>
      <guid>https://dev.to/jeffreyturov/i-built-a-free-aeo-checker-is-your-website-visible-to-chatgpt-perplexity-and-google-ai-10gl</guid>
      <description>&lt;p&gt;A restaurant owner in my town asked me a question I hear more and more: &lt;em&gt;"When someone asks ChatGPT where to eat tonight, does it say my name?"&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Honest answer: nobody knows without checking. So I built a free tool that checks — and I open-sourced the logic as an Apify Actor anyone (human or AI agent) can call.&lt;/p&gt;

&lt;h2&gt;
  
  
  The 7 things AI engines actually check
&lt;/h2&gt;

&lt;p&gt;After auditing dozens of local business sites, the signal for "will an AI cite you" collapses to 7 technical criteria:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Structured data (JSON-LD)&lt;/strong&gt; — machines read schema.org, not your hero banner&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A business schema type&lt;/strong&gt; — &lt;code&gt;LocalBusiness&lt;/code&gt;, &lt;code&gt;Organization&lt;/code&gt;, &lt;code&gt;ProfessionalService&lt;/code&gt;…&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A meta description&lt;/strong&gt; — the first text an AI reads about you&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One clear H1&lt;/strong&gt; — confirms the page's subject&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Open Graph tags&lt;/strong&gt; — sharing + entity understanding&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A &lt;code&gt;/llms.txt&lt;/code&gt; file&lt;/strong&gt; — the identity card AI engines look for (most sites don't have one)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI bots allowed in robots.txt&lt;/strong&gt; — most sites &lt;em&gt;accidentally&lt;/em&gt; block GPTBot, ClaudeBot or PerplexityBot via a blanket rule or a "SEO tool" that added it silently&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The killer is #7. I've seen sites with perfect schema markup that explicitly tell every AI crawler to go away.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the Actor returns
&lt;/h2&gt;

&lt;p&gt;One call, one score, exact fixes:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://www.codeur.com"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"scoreLabel"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"4/7"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"checks"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"json_ld"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"schema_type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"meta_description"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"h1"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"open_graph"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"llms_txt"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"robots_ai"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"topFixes"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"criterion"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Structured data (JSON-LD)"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"fix"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Add a &amp;lt;script type=&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;application/ld+json&lt;/span&gt;&lt;span class="se"&gt;\"&lt;/span&gt;&lt;span class="s2"&gt;&amp;gt; block..."&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="nl"&gt;"criterion"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"llms.txt file"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"fix"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Publish /llms.txt summarizing who you are for AI engines."&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;(That's a real result — a major marketplace scores 4/7. Most local businesses score 1–2.)&lt;/p&gt;

&lt;p&gt;It can also generate a &lt;strong&gt;ready-to-publish &lt;code&gt;llms.txt&lt;/code&gt; draft&lt;/strong&gt; for the audited site, delivered as plain text in the key-value store.&lt;/p&gt;

&lt;h2&gt;
  
  
  Call it from anything
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-X&lt;/span&gt; POST &lt;span class="s2"&gt;"https://api.apify.com/v2/acts/travelmonitorlab~aeo-ai-visibility-checker/run-sync-get-dataset-items?token=&lt;/span&gt;&lt;span class="nv"&gt;$APIFY_TOKEN&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{"url": "https://your-site.com", "generateLlmsTxt": true}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It's API-first and works through the Apify MCP server, so an LLM agent can audit a website mid-conversation. No browser, no proxy — a plain HTTP audit that runs in seconds and costs nothing. It's &lt;strong&gt;free&lt;/strong&gt;, deliberately: it's my 12th Actor and the first one I give away, because an AEO checker is exactly the kind of tool that should be a utility, not a product.&lt;/p&gt;

&lt;h2&gt;
  
  
  The bug that taught me something
&lt;/h2&gt;

&lt;p&gt;First deploy: every run returned &lt;code&gt;SUCCEEDED&lt;/code&gt;… with an empty dataset and a silent log. The cause wasn't the platform — it was me: I defined &lt;code&gt;async def main()&lt;/code&gt; and never called it. The module loaded, defined functions, and exited 0. One missing line:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;__name__&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;__main__&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;asyncio&lt;/span&gt;
    &lt;span class="n"&gt;asyncio&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;main&lt;/span&gt;&lt;span class="p"&gt;())&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A "successful" run proves nothing. Check the dataset, always.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Actor (free): &lt;a href="https://apify.com/travelmonitorlab/aeo-ai-visibility-checker" rel="noopener noreferrer"&gt;https://apify.com/travelmonitorlab/aeo-ai-visibility-checker&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;I also wrote a plain-language guide for non-devs: &lt;a href="https://adagasolutions.lu/ressources/etre-cite-par-chatgpt-commerce-local.html" rel="noopener noreferrer"&gt;Comment être cité par ChatGPT quand on est un commerce local&lt;/a&gt; (French, with the same 7 criteria explained)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you run it on your site, drop your score in the comments — I'm curious what the distribution looks like.&lt;/p&gt;

</description>
      <category>aeo</category>
      <category>seo</category>
      <category>ai</category>
      <category>apify</category>
    </item>
    <item>
      <title>Nature Detective — the offline AI that sends my kids back outside</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Tue, 06 Oct 2026 09:10:54 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/nature-detective-the-offline-ai-that-sends-my-kids-back-outside-5e5g</link>
      <guid>https://dev.to/jeffreyturov/nature-detective-the-offline-ai-that-sends-my-kids-back-outside-5e5g</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week1-2026-10-05"&gt;Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Last Sunday I watched my kids walk past a hundred fascinating things without seeing a single one.&lt;/p&gt;

&lt;p&gt;We live in Mersch, in the middle of Luxembourg — fifteen minutes from forests that have more species per square meter than my phone has apps. And yet the walk was negotiations: five more minutes of screen, then we'll go. The screen is the destination. The forest is the commute.&lt;/p&gt;

&lt;p&gt;So for this challenge I built the opposite deal: &lt;strong&gt;the screen becomes the reason to go into the forest — and the forest stays the point.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Nature Detective&lt;/strong&gt; is an offline-first nature companion for kids aged 5-10.&lt;/p&gt;

&lt;p&gt;A child photographs something alive — a plant, a beetle, a mushroom. An &lt;strong&gt;open-weight vision model (Gemma 3 4B, running entirely on local hardware)&lt;/strong&gt; identifies the find, tells one true fun fact a seven-year-old would love, sets a safety rule, and then does the thing no engagement-optimized app would ever do: &lt;strong&gt;it sends the kid back outside.&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;🎯 &lt;em&gt;"Count how many different colored flowers you can spot before the next corner!"&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Every identification comes with a mission — count, find, compare, observe. Missions never involve the screen. The screen is the shortest part of the experience, by design.&lt;/p&gt;

&lt;p&gt;Every discovery lands in a &lt;strong&gt;field journal&lt;/strong&gt; on the device, with a detective rank that grows with real outdoor finds (🐣 → 🐾 → 🦊 → 🦉). No account, no ads, no streak anxiety — just a collection of afternoons.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://main-michael-stanford-myrtle.trycloudflare.com" rel="noopener noreferrer"&gt;Try the live demo&lt;/a&gt;&lt;/strong&gt; — it runs on my own machine, on Gemma 3, right now.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fum3qhout3dvhrby56zxp.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fum3qhout3dvhrby56zxp.png" alt="The result screen: a bee identified, a safety badge, the next outdoor mission" width="800" height="2089"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;The full flow, no signup, works from a phone browser:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Open the &lt;a href="https://main-michael-stanford-myrtle.trycloudflare.com" rel="noopener noreferrer"&gt;live demo&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Snap a discovery (or upload any plant/insect photo)&lt;/li&gt;
&lt;li&gt;Get the identification, the fact, the safety badge — and your mission&lt;/li&gt;
&lt;li&gt;Check the field journal&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Fair warning: inference on a 24-core CPU takes ~15-60 seconds depending on load. Slow enough to look at the real thing while you wait — which, honestly, is also the point.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/jeffreyturov-dev" rel="noopener noreferrer"&gt;
        jeffreyturov-dev
      &lt;/a&gt; / &lt;a href="https://github.com/jeffreyturov-dev/nature-detective" rel="noopener noreferrer"&gt;
        nature-detective
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;🔍 Nature Detective&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;An offline-first nature companion that gets kids off the screen and into the woods.&lt;/strong&gt;
Built for the Hacktoberfest 2026 DEV Challenge — Week 1: &lt;em&gt;Touch Grass&lt;/em&gt;.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Live demo:&lt;/strong&gt; &lt;a href="https://main-michael-stanford-myrtle.trycloudflare.com" rel="nofollow noopener noreferrer"&gt;https://main-michael-stanford-myrtle.trycloudflare.com&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;A child on a hike photographs a plant, a bug, a mushroom. An &lt;strong&gt;open-weight vision
model (Gemma 3 4B) running locally&lt;/strong&gt; identifies the find, tells one true fun fact
sets a safety rule, and hands out the &lt;strong&gt;next outdoor mission&lt;/strong&gt; — count, find
compare. Every discovery lands in a field journal on the device.&lt;/p&gt;
&lt;p&gt;No account. No cloud. No photo, no GPS coordinate, no child's data ever leaves
your own hardware. It runs where the forest has no bars — that's the whole point.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Why open matters here&lt;/h2&gt;

&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;No internet, no problem.&lt;/strong&gt; Inference runs on a local machine (a laptop, or a
Raspberry Pi in a backpack acting as a Wi-Fi hotspot). The forest has no 5G…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/jeffreyturov-dev/nature-detective" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The whole stack, MIT licensed:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Piece&lt;/th&gt;
&lt;th&gt;Choice&lt;/th&gt;
&lt;th&gt;Why&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Vision + reasoning&lt;/td&gt;
&lt;td&gt;Gemma 3 4B via Ollama (Q4_K_M)&lt;/td&gt;
&lt;td&gt;open-weight, multimodal, runs on CPU&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Backend&lt;/td&gt;
&lt;td&gt;FastAPI + SQLite&lt;/td&gt;
&lt;td&gt;one file, zero infra&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Frontend&lt;/td&gt;
&lt;td&gt;Vanilla JS PWA, zero external assets&lt;/td&gt;
&lt;td&gt;the offline claim has to be true&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cost&lt;/td&gt;
&lt;td&gt;€0&lt;/td&gt;
&lt;td&gt;no API key, no per-call fee, ever&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;One model call turns a photo into a strict JSON object: identification, kid fact, safety level, next mission, quiz question. Everything the screen shows, in a single pass — because a children's app in a forest cannot afford round trips, and mine literally has none to afford: it runs where there is no signal.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;📱 phone camera ──► FastAPI ──► Gemma 3 (local vision) ──► strict JSON
                                                              │
        field journal (SQLite) ◄── safety net (code) ◄── verifier pass
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The interesting part is not the pipeline. It's what the pipeline refuses to do.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part I'm most proud of: what it refuses to do
&lt;/h2&gt;

&lt;p&gt;During testing, I fed the app an abstract watercolor — green and yellow blur, no living thing in it. A single-pass version of the app answered, with high confidence: &lt;strong&gt;"Spiderweb, Araneae."&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;It was a confident, detailed, completely fabricated answer. The most dangerous kind of wrong — the kind a child would believe, remember, and repeat at school.&lt;/p&gt;

&lt;p&gt;An app that teaches nature to kids has one job that outranks every other: &lt;strong&gt;never teach a lie.&lt;/strong&gt; So now every identification goes through a second pass — a verifier, running on the same local model, whose only job is to doubt the first answer. The trick that makes it actually work: the verifier &lt;strong&gt;never sees the claim first&lt;/strong&gt;. It must describe what it objectively sees — &lt;em&gt;"The image is an abstract painting, not a living organism."&lt;/em&gt; — and only then decide whether the claim matches its own independent observation. Ask it to judge the claim directly and it politely agrees with everything; make it commit to its own eyes first, and it catches the lie.&lt;/p&gt;

&lt;p&gt;The abstract painting now gets the honest answer: &lt;em&gt;"A first guess said this might be a spiderweb, but my double-check disagreed — and a good detective never teaches a maybe as a fact. Ask a grown-up or a field guide!"&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Teaching a child that "I don't know" is a respectable answer might be the most valuable feature in the whole app.&lt;/p&gt;

&lt;p&gt;Safety works the same way — enforced twice, because prompt rules are suggestions and &lt;strong&gt;code rules are guarantees&lt;/strong&gt;. The model is instructed that mushrooms and berries are always &lt;em&gt;look, don't touch&lt;/em&gt;. And then the backend overrides it in code anyway, every time, no matter what the model says. When I tested it with a fly agaric (&lt;em&gt;Amanita muscaria&lt;/em&gt; — the red one with white dots, beautiful and toxic), it identified it correctly &lt;em&gt;and&lt;/em&gt; refused any interaction with it, twice over.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;This project doesn't just &lt;em&gt;use&lt;/em&gt; open weights. &lt;strong&gt;It only exists because of them.&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The forest has no bars.&lt;/strong&gt; The whole point is a place with no signal. A closed API is a product that stops working exactly where mine starts working. Local inference isn't an optimization here — it's the product.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I will never upload my kids' photos to someone else's server.&lt;/strong&gt; Not their faces, not their location, not their finds. With Gemma running on my own hardware, nothing ever leaves the device. That's not a privacy policy; it's physics.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;€0, forever.&lt;/strong&gt; No API key, no subscription, no per-call cost — so it can run in a school, a scout group, a nature club with no budget, anywhere in the world.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ownable and swappable.&lt;/strong&gt; Any open-weight vision model Ollama can serve drops in with a one-line change. When a better small multimodal model ships next month, every Nature Detective gets smarter for free. Try getting that guarantee from a deprecated API version.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A closed model would have made this easier to demo and impossible to believe in.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma&lt;/strong&gt; — Gemma 3 4B is the entire brain of the project: species identification, kid-level explanations, mission generation, and the self-verification pass, all running locally through Ollama.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;This Sunday we're going back to the forest. My kids already asked if the detective is coming.&lt;/p&gt;

&lt;p&gt;He's in my pocket. He knows nothing about engagement metrics. And his favorite sentence, by far, is: &lt;em&gt;"Now go look for yourself."&lt;/em&gt; 🌿&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>opensource</category>
    </item>
    <item>
      <title>I Put 6,522 Birds in My Backpack: Offline Bird Call ID on a Laptop, Zero Internet</title>
      <dc:creator>Jeffrey Turov</dc:creator>
      <pubDate>Tue, 06 Oct 2026 08:11:43 +0000</pubDate>
      <link>https://dev.to/jeffreyturov/i-put-6522-birds-in-my-backpack-offline-bird-call-id-on-a-laptop-zero-internet-4pda</link>
      <guid>https://dev.to/jeffreyturov/i-put-6522-birds-in-my-backpack-offline-bird-call-id-on-a-laptop-zero-internet-4pda</guid>
      <description>&lt;h2&gt;
  
  
  The problem with birding apps
&lt;/h2&gt;

&lt;p&gt;The best birding happens exactly where your phone becomes a brick: deep forest, mountain trails, that marsh 40 minutes from the nearest cell tower. The moment a call you don't recognize echoes through the trees, the usual flow is: record it, &lt;em&gt;hope you remember it later&lt;/em&gt;, upload it when you get home, wait for the cloud.&lt;/p&gt;

&lt;p&gt;I wanted the whole loop to close &lt;strong&gt;on the trail&lt;/strong&gt;. So I built &lt;code&gt;trail-bird-id&lt;/code&gt;: point it at an audio file, get the species. No internet. No account. No API key. No per-call cost. Just a laptop (or a Raspberry Pi) and an open-weight model.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;$ &lt;/span&gt;python identify.py trail_recording_07h42.mp3

Analyzing trail_recording_07h42.mp3 &lt;span class="o"&gt;(&lt;/span&gt;fully offline, BirdNET V2.4, 6,522 species&lt;span class="o"&gt;)&lt;/span&gt;...

  1. Black-capped Sparrow &lt;span class="o"&gt;(&lt;/span&gt;Arremon abeillei&lt;span class="o"&gt;)&lt;/span&gt;      &lt;span class="nv"&gt;conf&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;0.99  @ 0.0-3.0s
  2. Streaked Saltator &lt;span class="o"&gt;(&lt;/span&gt;Saltator striatipectus&lt;span class="o"&gt;)&lt;/span&gt;   &lt;span class="nv"&gt;conf&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;0.86  @ 3.0-6.0s
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  The build: embarrassingly little code
&lt;/h2&gt;

&lt;p&gt;The heavy lifting is &lt;a href="https://github.com/birdnet-team/BirdNET-Analyzer" rel="noopener noreferrer"&gt;BirdNET&lt;/a&gt; — an open-source audio classifier from the Cornell Lab of Ornithology and Chemnitz University of Technology (CC BY-SA 4.0), trained on 6,522 species. It ships as a ~50 MB TFLite/TensorFlow model &lt;em&gt;inside the pip package&lt;/em&gt;, which is the whole trick: installing the library means installing the brain.&lt;/p&gt;

&lt;p&gt;My contribution is a ~60-line wrapper, &lt;a href="https://github.com/jeffreyturov-dev/trail-bird-id/blob/master/identify.py" rel="noopener noreferrer"&gt;&lt;code&gt;identify.py&lt;/code&gt;&lt;/a&gt;, that runs the model in 3-second windows over any wav/mp3/ogg file and aggregates to the best-confidence detection per species:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;uv venv &lt;span class="nt"&gt;--python&lt;/span&gt; 3.11 .venv
uv pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;--python&lt;/span&gt; .venv/bin/python birdnet-analyzer
.venv/bin/python identify.py your_recording.mp3 &lt;span class="nt"&gt;--top&lt;/span&gt; 5
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's the entire setup. No Docker, no GPU, no cloud credentials.&lt;/p&gt;

&lt;h2&gt;
  
  
  Does it actually work? Verified against ground truth
&lt;/h2&gt;

&lt;p&gt;I don't trust demos that only show one cherry-picked clip, so I tested against four field recordings from Wikimedia Commons where the species is known from the recording metadata (xeno-canto / iNaturalist sourced):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Recording&lt;/th&gt;
&lt;th&gt;Expected species&lt;/th&gt;
&lt;th&gt;Top-1 prediction&lt;/th&gt;
&lt;th&gt;Confidence&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Black-capped Sparrow XC250490 (Niels Krabbe, CC BY-SA)&lt;/td&gt;
&lt;td&gt;&lt;em&gt;Arremon abeillei&lt;/em&gt;&lt;/td&gt;
&lt;td&gt;✅ &lt;strong&gt;Black-capped Sparrow&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;0.99&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Australian Magpie song (CC BY-SA)&lt;/td&gt;
&lt;td&gt;&lt;em&gt;Gymnorhina tibicen&lt;/em&gt;&lt;/td&gt;
&lt;td&gt;✅ &lt;strong&gt;Australian Magpie&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;0.99&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Brown Hawk-Owl, South Bengal (CC BY)&lt;/td&gt;
&lt;td&gt;&lt;em&gt;Ninox scutulata&lt;/em&gt;&lt;/td&gt;
&lt;td&gt;✅ &lt;strong&gt;Brown Boobook&lt;/strong&gt; &lt;em&gt;(same bird, name updated by taxonomists — the model is more current than the file title)&lt;/em&gt;
&lt;/td&gt;
&lt;td&gt;1.00&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Guira Cuckoo (CC0)&lt;/td&gt;
&lt;td&gt;&lt;em&gt;Guira guira&lt;/em&gt;&lt;/td&gt;
&lt;td&gt;✅ &lt;strong&gt;Guira Cuckoo&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;0.95&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;4/4 top-1 correct&lt;/strong&gt;, across four continents and four very different vocalizations (sparrow song, magpie caroling, an owl's hoot, cuckoo chatter). Runtime: ~6.7 seconds wall-clock for a 40-second recording, on a plain CPU.&lt;/p&gt;

&lt;h2&gt;
  
  
  Proving "offline" isn't marketing
&lt;/h2&gt;

&lt;p&gt;"Works offline" is easy to claim. So I re-ran identification with &lt;strong&gt;all egress blocked&lt;/strong&gt; — every proxy environment variable pointed at a dead localhost port, so any HTTP call from the Python stack would fail instantly:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;$ &lt;/span&gt;&lt;span class="nb"&gt;env &lt;/span&gt;&lt;span class="nv"&gt;HTTPS_PROXY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;http://127.0.0.1:9 &lt;span class="nv"&gt;HTTP_PROXY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;http://127.0.0.1:9 &lt;span class="se"&gt;\&lt;/span&gt;
      &lt;span class="nv"&gt;ALL_PROXY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;socks5://127.0.0.1:9 &lt;span class="nv"&gt;NO_PROXY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
      .venv/bin/python identify.py audio/australian_magpie.ogg

  1. Australian Magpie &lt;span class="o"&gt;(&lt;/span&gt;Gymnorhina tibicen&lt;span class="o"&gt;)&lt;/span&gt;   &lt;span class="nv"&gt;conf&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;0.99  @ 9.0-12.0s
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Still works, because there is no network code to fail. The weights live in &lt;code&gt;site-packages/birdnet_analyzer/checkpoints/&lt;/code&gt;. The trail is the deployment target, and the trail has no SLA.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why open innovation matters here
&lt;/h2&gt;

&lt;p&gt;This is where the open approach doesn't just match the closed one — it &lt;em&gt;wins&lt;/em&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;It works where the closed apps can't.&lt;/strong&gt; Zero-connectivity is the core use case, not an edge case. A cloud API is a non-starter on a ridgeline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Nothing leaves your device.&lt;/strong&gt; Field recordings often capture more than birds — your voice, your location, your hiking companions. With a local model, that data never touches a server you don't control.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;You can actually own it.&lt;/strong&gt; Pin the model version, tune the confidence threshold for your region, swap in a species list for your local patch (&lt;code&gt;--lat/--lon&lt;/code&gt; filters to species plausible for your coordinates), or fine-tune on your own recordings. Try doing that with a closed mobile app.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost at scale is zero.&lt;/strong&gt; A citizen-science project processing 10,000 hours of soundscape pays nothing per call — because there are no calls.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Honest limitations
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;BirdNET needs a reasonably clean recording; heavy wind or overlapping dawn-chorus can drop confidence. The 3-second windowing means very short calls can be missed.&lt;/li&gt;
&lt;li&gt;My test set is 4 recordings, not 4,000 — it's a smoke test with known ground truth, not a benchmark. (BirdNET's own peer-reviewed evaluations cover the rigorous part.)&lt;/li&gt;
&lt;li&gt;I built and validated this on real field recordings made by others; the full "take it outside" bonus round happens this weekend on my local trails — I'll update the repo with what it hears.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Links
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Repo:&lt;/strong&gt; &lt;a href="https://github.com/jeffreyturov-dev/trail-bird-id" rel="noopener noreferrer"&gt;https://github.com/jeffreyturov-dev/trail-bird-id&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; &lt;a href="https://github.com/birdnet-team/BirdNET-Analyzer" rel="noopener noreferrer"&gt;BirdNET-Analyzer&lt;/a&gt; (CC BY-SA 4.0)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The screen is the shortest part of this experience: record outside, identify anywhere, get back to listening. 🌲&lt;/p&gt;

</description>
      <category>touchgrass</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
