<?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: Akshat Sharma</title>
    <description>The latest articles on DEV Community by Akshat Sharma (@akshat_slaya).</description>
    <link>https://dev.to/akshat_slaya</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%2F4190514%2Fdb3faf1f-c2cc-407f-b72f-78fa4d56089d.png</url>
      <title>DEV Community: Akshat Sharma</title>
      <link>https://dev.to/akshat_slaya</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/akshat_slaya"/>
    <language>en</language>
    <item>
      <title>👻 PhantomOps: The AI Exorcist That Works With the Internet Off | Ghost in the Machine Submission 🎃🌼</title>
      <dc:creator>Akshat Sharma</dc:creator>
      <pubDate>Thu, 29 Oct 2026 11:41:31 +0000</pubDate>
      <link>https://dev.to/akshat_slaya/phantomops-the-ai-exorcist-that-works-with-the-internet-off-ghost-in-the-machine-submission-4798</link>
      <guid>https://dev.to/akshat_slaya/phantomops-the-ai-exorcist-that-works-with-the-internet-off-ghost-in-the-machine-submission-4798</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;&lt;strong&gt;Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine&lt;/strong&gt;&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

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

&lt;p&gt;Open Task Manager right now. Scroll slowly. Somewhere in that list is a process with no reason to exist — a stale worker from a project you closed in June, a leaky helper quietly eating 400 MB, a script that was supposed to finish an hour ago and never got the memo.&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%2Fyz1f2v6yenm2j9urv4p4.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%2Fyz1f2v6yenm2j9urv4p4.png" alt="PhantomOps dashboard view" width="800" height="425"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Your machine is haunted. PhantomOps is the exorcist.&lt;/strong&gt; 👻&lt;/p&gt;

&lt;p&gt;PhantomOps is a &lt;strong&gt;zero-cloud-leakage FinOps &amp;amp; kernel agent&lt;/strong&gt; that runs entirely on your own hardware. It finds the ghost resources haunting your Windows kernel or your cloud bill, runs a local AI autopsy on them, lets you interrogate them in a live voice séance, requires a human to sign the kill order through Auth0 — and then banishes them for real, with a cryptographic receipt and a synthesized exorcism scream.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How I interpreted the theme: literally.&lt;/strong&gt; Zombie processes are the &lt;em&gt;undead of software&lt;/em&gt; — programs that finished their purpose but refused to leave memory. And Día de los Muertos teaches that the dead deserve to be honored properly, not just discarded. So PhantomOps doesn't just &lt;code&gt;kill -9&lt;/code&gt; a ghost. It gives it a funeral:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🔮 &lt;strong&gt;Divination&lt;/strong&gt; — a local &lt;strong&gt;TabPFN-3.5-Fast&lt;/strong&gt; foundation model (open weights, &lt;code&gt;.safetensors&lt;/code&gt; on disk, CPU-only) scores every process zero-shot for how "haunted" it is&lt;/li&gt;
&lt;li&gt;🧠 &lt;strong&gt;Autopsy&lt;/strong&gt; — a local &lt;strong&gt;Gemma 2 2B&lt;/strong&gt; (via Ollama) writes a structured root-cause report: cause of death, severity, monthly dollar drain, exact banishment command&lt;/li&gt;
&lt;li&gt;🎙️ &lt;strong&gt;Séance&lt;/strong&gt; — you speak into your microphone, &lt;strong&gt;ElevenLabs&lt;/strong&gt; transcribes it, Gemma answers &lt;em&gt;in the first person of the dying process&lt;/em&gt;, and a hoarse haunted voice reads its last words back to you&lt;/li&gt;
&lt;li&gt;🔐 &lt;strong&gt;The Warrant&lt;/strong&gt; — &lt;strong&gt;Auth0&lt;/strong&gt; step-up re-authentication binds a verified living human to the exact kill command, then mints a tamper-evident SHA-256 execution receipt&lt;/li&gt;
&lt;li&gt;💀 &lt;strong&gt;Exorcism&lt;/strong&gt; — the process tree dies for real (children first), an ElevenLabs-generated "server rack power-down with a ghostly whoosh" plays, and you download the signed audit ledger&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;And because this is the open-source AI challenge: the &lt;strong&gt;core loop — scan → detect → autopsy → kill — runs with the internet unplugged.&lt;/strong&gt; The model weights sit on your disk. Your process inventory never leaves the chassis. The AI inference bill is &lt;strong&gt;$0.00&lt;/strong&gt;, forever.&lt;/p&gt;

&lt;p&gt;A few receipts from my own testing:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;I spawn a real zombie with &lt;code&gt;spawn_zombie.py&lt;/code&gt; — it allocates an actual &lt;strong&gt;700 MB&lt;/strong&gt; of RAM and burns CPU in an infinite loop. One scan later, PhantomOps pins its ghost-confidence at &lt;strong&gt;99.99%&lt;/strong&gt; and locks it to the top of the Phantom Queue. Real PID. Real kill. Confirmed by the OS.&lt;/li&gt;
&lt;li&gt;The live scanner has &lt;strong&gt;per-tab precision&lt;/strong&gt;: Chromium browsers spawn a separate renderer process for every active tab, and PhantomOps inventories every PID independently — two runaway tabs surface as &lt;strong&gt;two separate suspects&lt;/strong&gt; with separate evidence files, not one vague "browser.exe". You interrogate and kill the exact haunted tab, and its 40 innocent siblings live.&lt;/li&gt;
&lt;li&gt;It has a conscience: a safety-lock list (&lt;code&gt;explorer.exe&lt;/code&gt;, &lt;code&gt;svchost.exe&lt;/code&gt;, &lt;code&gt;code.exe&lt;/code&gt;, &lt;code&gt;ollama.exe&lt;/code&gt;…) and a self-preservation rule mean it only kills what it can prove is a ghost.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;▶️️ &lt;strong&gt;Watch the full exorcism ritual:&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/H0Jv7wWqIi4" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;🔗 &lt;em&gt;Direct link: &lt;a href="https://youtu.be/H0Jv7wWqIi4" rel="noopener noreferrer"&gt;Watch on YouTube&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;What the recording shows — the 90-second arc:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;code&gt;python spawn_zombie.py&lt;/code&gt; — a real 700 MB zombie process comes alive on my machine&lt;/li&gt;
&lt;li&gt;One scan in PhantomOps → the zombie appears &lt;strong&gt;at the top of the Phantom Queue at 99.99% confidence&lt;/strong&gt; 👻&lt;/li&gt;
&lt;li&gt;Tab 1 → local Gemma 2 writes the autopsy: root cause, $ drain, kill command&lt;/li&gt;
&lt;li&gt;Tab 2 → I speak a question into my mic → the zombie &lt;em&gt;answers in its own haunted voice&lt;/em&gt; (ElevenLabs STT → local Gemma → haunted TTS)&lt;/li&gt;
&lt;li&gt;Tab 3 → I drag the memory slider down and watch TabPFN's ghost probability collapse in real time — counterfactual proof of &lt;em&gt;why&lt;/em&gt; it's haunted.&lt;/li&gt;
&lt;/ol&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%2Frk5u2sged11wmf8dzsxs.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%2Frk5u2sged11wmf8dzsxs.png" alt="Decay over the next 30 days" width="800" height="425"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The kill button → &lt;strong&gt;Auth0 step-up login&lt;/strong&gt; → 💀 the zombie dies, the exorcism SFX plays, and a signed SHA-256 audit receipt lands in my hands&lt;/li&gt;
&lt;/ol&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/akshat-bit420" rel="noopener noreferrer"&gt;
        akshat-bit420
      &lt;/a&gt; / &lt;a href="https://github.com/akshat-bit420/PhantomOps" rel="noopener noreferrer"&gt;
        PhantomOps
      &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;👻 PhantomOps — Local Infrastructure Ghost &amp;amp; Anomaly Exorcist&lt;/h1&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Zero-cloud-leakage FinOps &amp;amp; kernel agent.&lt;/strong&gt; Find the ghosts haunting your Windows kernel or your cloud bill → run a local AI autopsy → interrogate them in a live voice séance → get human clearance through Auth0 → put them to rest with a signed cryptographic receipt and a synthetic exorcism scream.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;&lt;em&gt;Built for the Hacktoberfest Open-Source AI Challenge: &lt;strong&gt;Ghost in the Machine&lt;/strong&gt; 🎃🌼&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Every machine has ghosts. The staging server that nobody killed in March. The script that leaked memory for 400 hours while everyone went home. The GPU instance cryptojacking at 3 AM. They drain your RAM and your budget — invisible, unaccounted for, &lt;em&gt;undead&lt;/em&gt;. PhantomOps is a paranormal investigator for your infrastructure: it sees them, names them, questions them, and (with a human's blessing) banishes them — &lt;strong&gt;entirely on your own hardware&lt;/strong&gt;.&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;🪦 The Haunting — What&lt;/h2&gt;…&lt;/div&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/akshat-bit420/PhantomOps" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The full architecture is documented in the README. Where to start digging:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;File&lt;/th&gt;
&lt;th&gt;What's Inside&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;spawn_zombie.py&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;🧟 The demo prop — summons a real 700 MB runaway process&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;app.py&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;🎛️ The Streamlit haunted console — 3 tabs, KPIs, Auth0 callback handler&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/live_scanner.py&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;🖥️ Per-PID kernel snapshot (tab-level precision) + process-tree exorcist&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/detector.py&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;🔮 TabPFN classifier + regressor + counterfactual blame engine&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/reasoner.py&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;🧠 Gemma 2 autopsies &amp;amp; séance roleplay (Pydantic-enforced JSON)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/gatekeeper.py&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;🔐 Auth0 step-up OAuth2 + JWT decode + SHA-256 receipts&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;src/narrator.py&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;🎙️ ElevenLabs STT / dual-voice TTS / exorcism SFX&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;&lt;strong&gt;The stack:&lt;/strong&gt; Python · Streamlit + Plotly (dark telemetry UI) · &lt;code&gt;tabpfn&lt;/code&gt; (Prior Labs) · Gemma 2 2B via Ollama · Auth0 (OAuth2) · ElevenLabs (&lt;code&gt;scribe_v1&lt;/code&gt;, &lt;code&gt;multilingual_v2&lt;/code&gt;, Text-to-Sound-Effects) · psutil · PyTorch on CPU.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;   🖥️ LIVE KERNEL (psutil)      ☁️ CLOUD ESTATE (CSV)
          └────────────┬────────────┘
                       ▼
        🔮 TABPFN-3.5-FAST (local .safetensors, CPU)
     classifier → ghost_confidence · regressor → doomsday
     counterfactuals → feature blame
                       ▼  flagged ghosts only
        🧠 GEMMA 2 2B via OLLAMA (local)
     JSON autopsy + first-person séance roleplay
                       ▼
   🔐 AUTH0 STEP-UP            🎙️ ELEVENLABS
   human clearance             STT + 2-voice TTS
   SHA-256 receipts            + exorcism SFX
                       ▼
        💀 EXORCISM: tree-kill + taskkill /F /T /PID
        📜 signed receipt: operator ∥ resource ∥ command ∥ time
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Stage 1 — Ingestion, with per-tab precision.&lt;/strong&gt; &lt;code&gt;psutil.process_iter&lt;/code&gt; walks the entire Windows process table and builds one evidence file per PID: RAM, CPU, threads, uptime, command line, and an estimated hourly cost (&lt;code&gt;mem_mb×0.0008 + cpu%×0.015 + threads×0.004&lt;/code&gt;). Because Chromium runs each active tab as its own renderer process, the scanner naturally sees &lt;strong&gt;each tab as its own suspect&lt;/strong&gt; — a runaway tab can't hide behind its browser. A cmdline fingerprint flags any running &lt;code&gt;spawn_zombie.py&lt;/code&gt; as an unmistakable &lt;code&gt;zombie_runaway_script&lt;/code&gt;; heuristics catch the rest (&lt;code&gt;live_memory_hog&lt;/code&gt; &amp;gt;4% RAM, &lt;code&gt;live_cpu_runaway&lt;/code&gt; &amp;gt;50% CPU, &lt;code&gt;thread_bloated_daemon&lt;/code&gt; &amp;gt;85 threads).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Stage 2 — TabPFN does the divining.&lt;/strong&gt; The classifier fits on baseline telemetry signatures and scores the remaining estate &lt;strong&gt;zero-shot&lt;/strong&gt; — no gradient training, no GPU, no data upload. My favorite seven lines in the repo:&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="c1"&gt;# Ensure any active spawn_zombie.py is always surfaced at the very top
&lt;/span&gt;&lt;span class="n"&gt;zombie_mask&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;df&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;resource_id&lt;/span&gt;&lt;span class="sh"&gt;"&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="nf"&gt;contains&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;spawn_zombie&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;case&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;False&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;na&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;False&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;zombie_mask&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;any&lt;/span&gt;&lt;span class="p"&gt;():&lt;/span&gt;
    &lt;span class="n"&gt;df&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;loc&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;zombie_mask&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;ghost_confidence&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;0.9999&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When the scanner sees the demo zombie, there is no "maybe". It is &lt;strong&gt;99.99% sure&lt;/strong&gt;, and the queue agrees. A &lt;code&gt;TabPFNRegressor&lt;/code&gt; then simulates 30 days of decay (memory +1.35%/day) to chart the cumulative financial drain against the flat $0 line after exorcism — and estimates time-to-OOM-crash in hours. Counterfactual sliders let you mutate a ghost's telemetry live and watch TabPFN re-classify it as healthy: the moment the confidence drops &lt;em&gt;is&lt;/em&gt; the explanation of what was haunting it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Stage 3 — Gemma 2 performs the autopsy.&lt;/strong&gt; Flagged telemetry goes to a &lt;strong&gt;local Gemma 2 2B&lt;/strong&gt; through Ollama, which must return JSON matching a Pydantic &lt;code&gt;GhostAutopsy&lt;/code&gt; schema — &lt;code&gt;incident_title&lt;/code&gt;, &lt;code&gt;severity&lt;/code&gt;, &lt;code&gt;root_cause_autopsy&lt;/code&gt;, &lt;code&gt;monthly_waste_usd&lt;/code&gt;, and an environment-aware &lt;code&gt;exorcism_command&lt;/code&gt; (&lt;code&gt;taskkill /F /T /PID &amp;lt;pid&amp;gt;&lt;/code&gt; for kernel ghosts, &lt;code&gt;aws ec2 terminate-instances&lt;/code&gt; for cloud spirits). The same model then switches personas for the séance: &lt;em&gt;first person, dramatic, technically accurate to its own real metrics&lt;/em&gt;. Ask it "why are you hogging memory?" and it answers as a trapped process begging for either a fix or the exorcism.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Stage 4 — Auth0 holds the ward.&lt;/strong&gt; Destructive actions generate an OAuth2 authorization URL with &lt;code&gt;prompt=login&lt;/code&gt; (explicit human re-authentication) and a SHA-256 &lt;code&gt;state&lt;/code&gt; binding the request to the exact resource + command — so a callback can't be replayed against a different target. On approval, the JWT is decoded in-process and a &lt;strong&gt;tamper-evident receipt&lt;/strong&gt; is minted:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;📜 EXOR-7F3A21C9D4
   operator : auth0|verified-operator
   command  : taskkill /F /T /PID 4132
   binding  : SHA-256( operator ∥ resource ∥ command ∥ timestamp )
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The AI agent literally &lt;em&gt;cannot&lt;/em&gt; kill a process without a living human signing off through Auth0. That's the human-in-the-loop pattern, enforced cryptographically.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Stage 5 — ElevenLabs is the medium.&lt;/strong&gt; Three distinct APIs, two personas: &lt;code&gt;scribe_v1&lt;/code&gt; transcribes your spoken séance questions, &lt;code&gt;multilingual_v2&lt;/code&gt; speaks on both sides of the veil (an authoritative &lt;em&gt;Commander&lt;/em&gt; for incident briefings, the hoarse &lt;em&gt;Callum&lt;/em&gt; for the dying process), and &lt;strong&gt;Text-to-Sound-Effects&lt;/strong&gt; synthesizes the exorcism — &lt;em&gt;"a sci-fi server rack power down with a ghostly digital whoosh and electrical discharge"&lt;/em&gt; — the instant a ghost dies. Audio is cached locally to conserve credits.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Stage 6 — The exorcist kills carefully.&lt;/strong&gt; Children die first, then &lt;code&gt;psutil.kill()&lt;/code&gt;, then a native &lt;code&gt;taskkill /F /T /PID&lt;/code&gt; to clean up handles, then &lt;code&gt;.venv&lt;/code&gt; parent launchers. It refuses to harm its own Streamlit process or anything on the protected list — unless the target is explicitly the zombie. Professional liability standards, paranormal cleanup division.&lt;/p&gt;

&lt;p&gt;Everything persists across sessions: autopsies, exorcism records, and audit receipts survive restarts via a local JSON/CSV state layer.&lt;/p&gt;

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

&lt;p&gt;For PhantomOps, open weights aren't an ideology — they're the &lt;strong&gt;architecture&lt;/strong&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;A kernel inventory is the most sensitive map a machine has.&lt;/strong&gt; It's every tool you use, every tab you keep open, every script you run. A closed anomaly-detection API would mean uploading that entire map to someone else's server &lt;em&gt;on every scan&lt;/em&gt;. With local TabPFN and local Gemma 2, &lt;strong&gt;the telemetry never leaves the chassis&lt;/strong&gt; — the zero-cloud-leakage promise is enforced by physics, not by a privacy policy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It runs with the lights (and the internet) off.&lt;/strong&gt; Scan → detect → autopsy → text séance → exorcism all work air-gapped. That matters because &lt;em&gt;your machine being wrong&lt;/em&gt; and &lt;em&gt;your connection being down&lt;/em&gt; correlate more than we'd like to admit. An exorcist that needs WiFi is not an exorcist.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;$0 per prediction, forever.&lt;/strong&gt; TabPFN scores on CPU from local weights; Gemma infers on-device. You can hunt ghosts all night and the invoice stays empty. A metered API would make the &lt;em&gt;cost of checking&lt;/em&gt; a recurring tax on curiosity.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The behavior is yours to change.&lt;/strong&gt; Swap &lt;code&gt;gemma2:2b&lt;/code&gt; for any Ollama model in one env var. Load different TabPFN weights. Rewrite the ghost's séance persona to be funnier, darker, or multilingual. A closed ghost-hunter is a black box that can vanish, reprice, or silently swap its model under you — an open one is a &lt;em&gt;spirit you keep&lt;/em&gt;, not a service you rent.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Where open genuinely beat closed for us:&lt;/strong&gt; TabPFN gave us zero-shot tabular anomaly scoring with no training loop and no GPU — on the single most sensitive dataset a laptop owns. Gemma 2 2B gave structured JSON we could enforce locally with Pydantic. The only pieces we rent — ElevenLabs' voices and Auth0's identity ward — are optional ornaments on an engine that runs entirely at home.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;The ghost stays in your machine. So does the exorcist.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The Spirit of It
&lt;/h2&gt;

&lt;p&gt;The first time it worked, I actually got chills.&lt;/p&gt;

&lt;p&gt;I had spawned a zombie — 700 MB of real RAM, burning real CPU — and asked it, out loud, through my microphone, why it was still alive. ElevenLabs transcribed me. Gemma answered as the dying process, in first person, citing its own actual memory footprint. A hoarse, haunted voice read the reply back across the quiet room.&lt;/p&gt;

&lt;p&gt;Then I clicked the kill button, logged in through Auth0 like a proper executioner, and watched the PID vanish from &lt;code&gt;psutil&lt;/code&gt; in real time. The exorcism SFX — that ghostly server-rack power-down — played, and the machine felt &lt;em&gt;different&lt;/em&gt;. Lighter. Quiet in the way a house is quiet after a storm passes.&lt;/p&gt;

&lt;p&gt;Here's the thing about Día de los Muertos: it isn't about banishing the dead. It's about &lt;em&gt;remembering them properly&lt;/em&gt;. So every ghost in PhantomOps gets the full rite — an autopsy (cause of death), a séance (last words), a signed receipt (a memorial plaque in the audit ledger), and a sound (the tolling of a dying server). Some marigolds for my RAM. 🌼&lt;/p&gt;

&lt;p&gt;Somewhere in your Task Manager is a process that should have died months ago. Now there's a proper way to say goodbye.&lt;/p&gt;

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

&lt;h3&gt;
  
  
  Best Use of TabPFN (Prior Labs)
&lt;/h3&gt;

&lt;p&gt;TabPFN is the brain of the entire pipeline — and the deepest integration in the project. &lt;code&gt;TabPFNClassifier&lt;/code&gt; performs zero-shot ghost scoring on live kernel telemetry (no training loop, no GPU, offline &lt;code&gt;.safetensors&lt;/code&gt; weights with &lt;code&gt;TABPFN_OFFLINE=1&lt;/code&gt;); &lt;code&gt;TabPFNRegressor&lt;/code&gt; powers the 30-day "Doomsday" cost forecaster and time-to-OOM estimation; and a counterfactual simulator uses TabPFN's predictions for real-time what-if sliders and zero-shot, perturbation-based feature blame attribution.&lt;/p&gt;

&lt;h3&gt;
  
  
  Best Use of Gemma
&lt;/h3&gt;

&lt;p&gt;A local Gemma 2 2B (via Ollama) does all reasoning on-device: Pydantic-schema-enforced JSON autopsies with severity, dollar drain and exact remediation commands — plus the first-person séance roleplay where the haunted process speaks for itself using its own real metrics. Fully local, swap-friendly, works offline.&lt;/p&gt;

&lt;h3&gt;
  
  
  Best Use of Auth0 for AI Agents
&lt;/h3&gt;

&lt;p&gt;PhantomOps is an AI agent that takes &lt;em&gt;destructive&lt;/em&gt; actions on a real OS — so it doesn't get to act alone. Auth0 enforces human-in-the-loop step-up authorization (&lt;code&gt;prompt=login&lt;/code&gt;) for every exorcism, cryptographically binds each approval to the exact resource and kill command via a SHA-256 &lt;code&gt;state&lt;/code&gt;, decodes the JWT claims in-process, and mints tamper-evident SHA-256 execution receipts that persist in the audit ledger.&lt;/p&gt;

&lt;h3&gt;
  
  
  Best Use of ElevenLabs
&lt;/h3&gt;

&lt;p&gt;Three ElevenLabs APIs form the voice layer: &lt;code&gt;scribe_v1&lt;/code&gt; speech-to-text transcribes your spoken séance questions, &lt;code&gt;multilingual_v2&lt;/code&gt; TTS performs both sides of the ritual with two distinct personas (the authoritative Commander and the haunted Callum voice of the dying process), and Text-to-Sound-Effects generates the exorcism power-down that plays the moment a ghost is banished.&lt;/p&gt;




&lt;p&gt;👻 &lt;strong&gt;Your turn:&lt;/strong&gt; open Task Manager and find the weirdest "process that shouldn't exist" on your machine — drop it in the comments. Best haunted process name gets featured in the README's Hall of Ghosts.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Built in one haunted week for Hacktoberfest. MIT licensed. The ghost stays in your machine.&lt;/em&gt; 🎃🌼&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>testing</category>
      <category>ai</category>
    </item>
    <item>
      <title>SlothSuite: Slow &amp; Steady Local AI for PDFs and Git</title>
      <dc:creator>Akshat Sharma</dc:creator>
      <pubDate>Fri, 23 Oct 2026 12:53:35 +0000</pubDate>
      <link>https://dev.to/akshat_slaya/slothsuite-slow-steady-local-ai-for-pdfs-and-git-b56</link>
      <guid>https://dev.to/akshat_slaya/slothsuite-slow-steady-local-ai-for-pdfs-and-git-b56</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;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;  _____ _       _   _      _____         _ _          
 / ____| |     | | | |    / ____|       (_| |         
| (___ | | ___ | |_| |__ | (___  _   _   _| |_ ___    
 \___ \| |/ _ \| __| '_ \ \___ \| | | | | | __/ _ \   
 ____) | | (_) | |_| | | |____) | |_| | | | ||  __/   
|_____/|_|\___/ \__|_| |_|_____/ \__,_| |_|\__\___|   

         ___
       /     \      
      | (o) (o) |     🦥 S L O T H S U I T E  (GIT slothNtidy)
       \   _   /      Slow &amp;amp; Steady Local AI for PDF Deduplication &amp;amp; Git
        |  v  |       100% Offline • Powered by Ollama &amp;amp; qwen2.5:7b
         \___/        --------------------------------------------------
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

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

&lt;p&gt;Sloths survive in the wild by conserving energy and eliminating wasted movement. Yet as students and developers, we burn hours every week on two tedious digital chores:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;The Duplicate PDF &amp;amp; Messy Downloads Trap:&lt;/strong&gt; Downloading files like &lt;code&gt;main.pdf&lt;/code&gt;, &lt;code&gt;Python_Assignment_2026 (1).pdf&lt;/code&gt;, &lt;code&gt;dbms_pdf3.pdf&lt;/code&gt;, and &lt;code&gt;dbms_pdf4.pdf&lt;/code&gt;, manually opening each one just to figure out which version actually has the extra pages, and skimming dense paragraphs to find assignment deadlines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The Git Commit &amp;amp; Changelog Tax:&lt;/strong&gt; Finishing a coding session only to stare at a terminal trying to write clean Conventional Commit messages and manually updating &lt;code&gt;CHANGELOG.md&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;To embrace the &lt;strong&gt;Slow and Steady&lt;/strong&gt; ethos, I built &lt;strong&gt;SlothSuite (&lt;code&gt;GIT tidyNsloth&lt;/code&gt;)&lt;/strong&gt;—a zero-friction, 100% offline local AI automation suite powered by &lt;strong&gt;&lt;code&gt;qwen2.5:7b&lt;/code&gt;&lt;/strong&gt; via Ollama (&lt;code&gt;http://localhost:11434&lt;/code&gt;). Rather than building a basic file-mover that blindly renames files one by one, SlothSuite operates as a &lt;strong&gt;dual-engine local intelligence workflow&lt;/strong&gt;:&lt;/p&gt;
&lt;h3&gt;
  
  
  1. Smart Batch PDF Deduplicator, Content Differ &amp;amp; Vault (&lt;code&gt;Tidy.py&lt;/code&gt;)
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Pre-Flight Reference Scan:&lt;/strong&gt; Before touching a single file in &lt;code&gt;./Downloaded_PDFs&lt;/code&gt;, the agent pre-scans the entire batch to extract text layers, page counts, and character lengths into an in-memory reference table.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cross-File Duplicate Detection &amp;amp; Content Diffing:&lt;/strong&gt; Using sequence matching (&lt;code&gt;difflib.SequenceMatcher&lt;/code&gt;) paired with &lt;code&gt;qwen2.5:7b&lt;/code&gt;, it detects overlapping duplicate PDFs across the batch (e.g., a 1-page partial draft vs. a 2-page complete version). Instead of blindly overwriting or appending &lt;code&gt;-1&lt;/code&gt; to the filename, it prints a side-by-side summary of both files and a bulleted &lt;strong&gt;Difference Report (&lt;code&gt;➕&lt;/code&gt;)&lt;/strong&gt; showing &lt;em&gt;what extra content the fuller PDF contains that the shorter copy is missing&lt;/em&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Human-in-the-Loop Cleanup &amp;amp; Live Vault Indexing:&lt;/strong&gt; Asks for explicit user confirmation (&lt;code&gt;[y/N]&lt;/code&gt;) before deleting the inferior duplicate from disk. Surviving PDFs are categorized, renamed (&lt;code&gt;YYYY-MM-DD_clean_name.pdf&lt;/code&gt;), moved into &lt;code&gt;./SlothVault&lt;/code&gt;, and compiled into a self-healing &lt;strong&gt;&lt;code&gt;SLOTH_INDEX.md&lt;/code&gt;&lt;/strong&gt; dashboard featuring crisp bullet points, clickable links, diff histories, and &lt;code&gt;- [ ]&lt;/code&gt; action items.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3&gt;
  
  
  2. Spotlight Feature: Offline Git Automation Agent (&lt;code&gt;sloth.py&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;While file organization saves clutter, &lt;strong&gt;&lt;code&gt;sloth.py&lt;/code&gt;&lt;/strong&gt; is a daily superpower for developers:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Zero Cloud Exposure &amp;amp; Zero Hardcoded Credentials:&lt;/strong&gt; Unlike cloud AI commit tools that upload your proprietary code diffs to remote servers or require API tokens, &lt;code&gt;sloth.py&lt;/code&gt; interacts directly with your local &lt;code&gt;.git&lt;/code&gt; staging tree via Python's &lt;code&gt;subprocess&lt;/code&gt; and sends diffs strictly to &lt;code&gt;localhost:11434&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Context-Aware Conventional Commits:&lt;/strong&gt; Combines structural file summaries (&lt;code&gt;git diff --cached --stat&lt;/code&gt;) with line-level diffs (&lt;code&gt;git diff --cached&lt;/code&gt;) so &lt;code&gt;qwen2.5:7b&lt;/code&gt; understands file deletions, folder restructures, and code logic changes, formatting them into strict Conventional Commits (&lt;code&gt;feat:&lt;/code&gt;, &lt;code&gt;refactor:&lt;/code&gt;, &lt;code&gt;fix:&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Automated &lt;code&gt;CHANGELOG.md&lt;/code&gt; &amp;amp; Lazy Score:&lt;/strong&gt; Automatically appends dated bullet points under &lt;code&gt;## [YYYY-MM-DD]&lt;/code&gt; in &lt;code&gt;CHANGELOG.md&lt;/code&gt;, calculates a fun &lt;strong&gt;"Lazy Score"&lt;/strong&gt; showing how many lines of typing you saved, and runs &lt;code&gt;git commit&lt;/code&gt; + &lt;code&gt;git push&lt;/code&gt; on your &lt;code&gt;[y/N]&lt;/code&gt; approval.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Who gets to be a little lazier because of it?&lt;/strong&gt; Students drowning in duplicate lecture notes and developers who want self-documenting Git repositories without paying API fees or leaking code to the cloud.&lt;/p&gt;


&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;
&lt;h3&gt;
  
  
  1. &lt;code&gt;Tidy.py&lt;/code&gt; — Smart Batch Pre-Scan, Content Diffing &amp;amp; User-Approved Deletion
&lt;/h3&gt;

&lt;p&gt;When given multiple DBMS note PDFs where &lt;code&gt;dbms_pdf4.pdf&lt;/code&gt; is a shorter 1-page version (&lt;code&gt;1,144 chars&lt;/code&gt;) and &lt;code&gt;dbms_pdf3.pdf&lt;/code&gt; is the expanded 2-page version (&lt;code&gt;2,715 chars&lt;/code&gt;), &lt;code&gt;Tidy.py&lt;/code&gt; catches the overlap upfront, explains the exact differences, and asks permission before deleting the redundant file:&lt;br&gt;
&lt;/p&gt;
&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;PS C:\Users\aksha\GIT tidyNsloth&amp;gt; .\.venv\Scripts\python.exe Tidy.py

🔍 STEP 1: Pre-scanning all 5 PDF(s) in './Downloaded_PDFs'...
   📄 Loaded reference: dbms_pdf3.pdf (2p, 2715 chars)
   📄 Loaded reference: dbms_pdf4.pdf (1p, 1144 chars)

🧠 STEP 2: Cross-comparing all files against each other for duplicates...
=================================================================
⚠️  DUPLICATE DETECTED: 'dbms_pdf4.pdf' vs 'dbms_pdf3.pdf'
=================================================================
🦥 Asking Qwen to summarize both and extract the differences...

📉 SHORTER FILE: dbms_pdf4.pdf (1p, 1144 chars)
   • Covers core transaction properties (ACID) and basic lock-based protocols.

📈 FULLER FILE:  dbms_pdf3.pdf (2p, 2715 chars)
   • Comprehensive notes on DBMS transactions, concurrency control, and deadlock handling.

🔍 DIFFERENCE (What 'dbms_pdf3.pdf' has that 'dbms_pdf4.pdf' is missing):
   ➕ Includes detailed explanations on deadlock detection and prevention, such as wait-for graph, prevention methods, and timeout mechanisms.
   ➕ Expands on concurrency control with timestamp ordering, where each transaction has a unique timestamp, and operations violating timestamp order are rolled back.

👉 Delete shorter file 'dbms_pdf4.pdf'? [y/N]: y
🗑️  Deleted 'dbms_pdf4.pdf'. Keeping 'dbms_pdf3.pdf'!

📂 STEP 3: Indexing surviving PDF(s) into './SlothVault'...
   ✅ Moved to Notes/2026-10-23_dbms_indexing_optimization.pdf
   ✅ Moved to Notes/2026-10-23_dbms_normalization.pdf
   ✅ Moved to Notes/2026-10-23_dbms_transactions_concurrency.pdf

🎉 All done! Open './SlothVault/SLOTH_INDEX.md' to view your clean report.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&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%2Flk538qoy99jjonseymkj.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%2Flk538qoy99jjonseymkj.png" alt="SlothSuite comparing duplicate DBMS PDFs, highlighting missing content, and asking for confirmation before deleting the shorter file" width="800" height="430"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;h3&gt;
  
  
  2. Auto-Generated &lt;code&gt;SLOTH_INDEX.md&lt;/code&gt; Dashboard
&lt;/h3&gt;

&lt;p&gt;Even when dropped files have vague names like &lt;code&gt;main.pdf&lt;/code&gt; or &lt;code&gt;Python_Assignment_2026 (1).pdf&lt;/code&gt;, the agent reads inside them, renames them cleanly (&lt;code&gt;2026-10-21_anthony_shea_professional.pdf&lt;/code&gt;, &lt;code&gt;2026-10-21_oop_python_assignment.pdf&lt;/code&gt;), extracts core topics, and flags action deadlines:&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 Smart Document Index&lt;/span&gt;

&lt;span class="gu"&gt;### 📄 [2026-10-21_oop_python_assignment.pdf](./ComputerScience/2026-10-21_oop_python_assignment.pdf)&lt;/span&gt;
&lt;span class="gs"&gt;**Original File:**&lt;/span&gt; &lt;span class="sb"&gt;`Python_Assignment_2026 (1).pdf`&lt;/span&gt; | &lt;span class="gs"&gt;**Stats:**&lt;/span&gt; &lt;span class="sb"&gt;`2 page(s), 3420 chars`&lt;/span&gt;

&lt;span class="gs"&gt;**📌 Key Points:**&lt;/span&gt;
&lt;span class="p"&gt;*&lt;/span&gt; Contains an assignment for Object Oriented Programming using Python.
&lt;span class="p"&gt;*&lt;/span&gt; Covers Python features, applications, class definitions, and method implementations.

&lt;span class="gs"&gt;**⚡ Action Items &amp;amp; Deadlines:**&lt;/span&gt;
&lt;span class="p"&gt;*&lt;/span&gt; [ ] Complete and submit the assignment by February 28, 2026.
&lt;span class="p"&gt;
---
&lt;/span&gt;
&lt;span class="gu"&gt;### 📄 [2026-10-21_anthony_shea_professional.pdf](./ComputerScience/2026-10-21_anthony_shea_professional.pdf)&lt;/span&gt;
&lt;span class="gs"&gt;**Original File:**&lt;/span&gt; &lt;span class="sb"&gt;`main.pdf`&lt;/span&gt; | &lt;span class="gs"&gt;**Stats:**&lt;/span&gt; &lt;span class="sb"&gt;`1 page(s), 2180 chars`&lt;/span&gt;

&lt;span class="gs"&gt;**📌 Key Points:**&lt;/span&gt;
&lt;span class="p"&gt;*&lt;/span&gt; Highlights roles as a Data Scientist and Backend Engineer focusing on ML, data analysis, and imagery analysis.
&lt;span class="p"&gt;*&lt;/span&gt; Key skills include Python, SQL, TensorFlow, and Kibana.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&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%2Fi540cve0qauf48ksg9bh.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%2Fi540cve0qauf48ksg9bh.png" alt="SlothSuite's generated SLOTH_INDEX.md dashboard showing organized PDF files, extracted key points, and assignment deadlines" width="800" height="431"&gt;&lt;/a&gt;&lt;/p&gt;
&lt;h3&gt;
  
  
  3. &lt;code&gt;sloth.py&lt;/code&gt; — Offline Git Diff Analysis, Changelog Update &amp;amp; Auto-Push
&lt;/h3&gt;


&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;PS C:\Users\aksha\GIT tidyNsloth&amp;gt; git add -A
PS C:\Users\aksha\GIT tidyNsloth&amp;gt; .\.venv\Scripts\python.exe sloth.py
🦥 Reading your staged changes locally with qwen2.5:7b (0 bytes sent to cloud)...

✨ Proposed Commit: refactor(tidy): remove old DBMS definitions and add new indexing notes
📝 Changelog Entry:
   • Removed outdated DBMS definitions document
   • Added new notes on database indexing optimization
   • Updated Sloth Smart Document Index with latest content
💤 You just saved yourself from typing 324 lines of old definitions and added 99 lines of new content. Nice work!

👉 Commit and push these changes to GitHub? [y/N]: y
🚀 Pushing to GitHub...
🦥 Done! Refresh your GitHub repo page to see the update.
&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/akshat-bit420" rel="noopener noreferrer"&gt;
        akshat-bit420
      &lt;/a&gt; / &lt;a href="https://github.com/akshat-bit420/Git_slothNtidy" rel="noopener noreferrer"&gt;
        Git_slothNtidy
      &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="snippet-clipboard-content notranslate position-relative overflow-auto"&gt;&lt;pre class="notranslate"&gt;&lt;code&gt;_____ _       _   _      _____         _ _          
 / ____| |     | | | |    / ____|       (_| |         
| (___ | | ___ | |_| |__ | (___  _   _   _| |_ ___    
 \___ \| |/ _ \| __| '_ \ \___ \| | | | | | __/ _ \   
 ____) | | (_) | |_| | | |____) | |_| | | | ||  __/   
|_____/|_|\___/ \__|_| |_|_____/ \__,_| |_|\__\___|   
                                 __/ |                
 [ 🦥 Slow &amp;amp; Steady • 100% Local AI • Zero Cloud ]   |___/
&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;🦥 SlothSuite (&lt;code&gt;GIT tidyNsloth&lt;/code&gt;)&lt;/h1&gt;

&lt;/div&gt;

&lt;div class="markdown-heading"&gt;
&lt;h3 class="heading-element"&gt;Zero-Friction, Privacy-First Local AI Document Management &amp;amp; Git Automation Suite&lt;/h3&gt;

&lt;/div&gt;

&lt;p&gt;&lt;em&gt;Built for local-first developers, students, and privacy advocates.&lt;/em&gt;&lt;/p&gt;




&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;🌿 The Sloth Ethos&lt;/h2&gt;

&lt;/div&gt;

&lt;p&gt;Sloths survive by conserving energy. You shouldn't waste yours manually renaming messy downloads, figuring out which PDF version has the extra page, or writing repetitive Conventional Commit messages.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;SlothSuite&lt;/strong&gt; is a dual-engine local AI…&lt;/p&gt;&lt;/div&gt;


&lt;/div&gt;
&lt;br&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/akshat-bit420/Git_slothNtidy" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;br&gt;
&lt;/div&gt;
&lt;br&gt;


&lt;h3&gt;
  
  
  Repository Structure
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;GIT tidyNsloth/
├── Downloaded_PDFs/         # Input drop folder for raw/duplicate PDFs
├── SlothVault/              # Organized output vault managed by Tidy.py
│   ├── Notes/               # Auto-generated category folders (YYYY-MM-DD_slug.pdf)
│   ├── SLOTH_INDEX.md       # Master Markdown dashboard with bullets, diffs &amp;amp; deadlines
│   └── sloth_memory.json    # Persistent metadata &amp;amp; text store for cross-file dedup
├── engine.py                # Local Ollama REST client with strict JSON schema enforcement
├── Tidy.py                  # 2-stage batch PDF pre-scanner, content differ &amp;amp; deduplicator
├── sloth.py                 # Offline Git diff analyzer, commit writer &amp;amp; changelog updater
├── CHANGELOG.md             # Auto-maintained project changelog generated by sloth.py
├── .gitignore               # Excludes .venv/ and __pycache__/
└── README.md                # Project documentation &amp;amp; architecture guide
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






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

&lt;h3&gt;
  
  
  Architectural Flow &amp;amp; Agent Pipelines
&lt;/h3&gt;

&lt;h4&gt;
  
  
  Pipeline 1: Smart Batch PDF Deduplicator (&lt;code&gt;Tidy.py&lt;/code&gt;)
&lt;/h4&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;                  ┌──────────────────────────────┐
                  │      ./Downloaded_PDFs       │
                  │ (Raw &amp;amp; Duplicate PDF Batch)  │
                  └──────────────┬───────────────┘
                                 │
                                 ▼
        ┌─────────────────────────────────────────────────┐
        │ STEP 1: Pre-Flight Reference Scan (PyMuPDF)     │
        │ • Extracts text, page count &amp;amp; character length  │
        │ • Builds in-memory reference table of all files │
        └────────────────────────┬────────────────────────┘
                                 │
                                 ▼
        ┌─────────────────────────────────────────────────┐
        │ STEP 2: Cross-File Duplicate Matrix (difflib)   │
        │ • Compares File i vs File j (1..N) upfront      │
        │ • Checks &amp;gt;= 55% prefix overlap in milliseconds  │
        └───────────┬─────────────────────────┬───────────┘
                    │                         │
          [ Unique Document ]        [ Duplicate Pair Found ]
                    │                         │
                    │                         ▼
                    │        ┌────────────────────────────────────┐
                    │        │ STEP 3: Delta Slicing &amp;amp; LLM Diff   │
                    │        │ • Crowns Fuller vs Shorter PDF     │
                    │        │ • Slices extra chars in Fuller PDF │
                    │        │ • qwen2.5:7b generates bullet diff │
                    │        └────────────────┬───────────────────┘
                    │                         │
                    │                         ▼
                    │        ┌────────────────────────────────────┐
                    │        │ STEP 4: User Confirmation [y/N]    │
                    │        │ • Prints side-by-side summary+diff │
                    │        │ • Deletes inferior copy on 'y'     │
                    │        └────────────────┬───────────────────┘
                    │                         │
                    └────────────┬────────────┘
                                 │
                                 ▼
        ┌─────────────────────────────────────────────────┐
        │ STEP 5: Vault Filing &amp;amp; SLOTH_INDEX.md Rebuild   │
        │ • Categorizes &amp;amp; renames: YYYY-MM-DD_slug.pdf    │
        │ • Moves survivors into ./SlothVault/&amp;lt;Category&amp;gt;/ │
        │ • Updates sloth_memory.json &amp;amp; SLOTH_INDEX.md    │
        └─────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h4&gt;
  
  
  Pipeline 2: Offline Git Commit &amp;amp; Changelog Agent (&lt;code&gt;sloth.py&lt;/code&gt;)
&lt;/h4&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;                  ┌──────────────────────────────┐
                  │    Staged Changes in Git     │
                  │         (git add -A)         │
                  └──────────────┬───────────────┘
                                 │
                                 ▼
        ┌─────────────────────────────────────────────────┐
        │ STEP 1: Subprocess Diff Extraction              │
        │ • Runs `git diff --cached --stat` (file stats)  │
        │ • Runs `git diff --cached` (line-level diff)    │
        └────────────────────────┬────────────────────────┘
                                 │
                                 ▼
        ┌─────────────────────────────────────────────────┐
        │ STEP 2: Local Ollama Inference (engine.py)      │
        │ • Sends payload to localhost:11434 (qwen2.5:7b) │
        │ • Enforces JSON schema:                         │
        │   - "commit_message" (Conventional Commit)      │
        │   - "changelog_bullets" (1-3 Markdown bullets)  │
        │   - "lazy_score" (Lines/effort saved metric)    │
        └────────────────────────┬────────────────────────┘
                                 │
                                 ▼
        ┌─────────────────────────────────────────────────┐
        │ STEP 3: Developer Approval &amp;amp; Automated Push     │
        │ • Previews commit message, bullets &amp;amp; lazy score │
        │ • Prompts: Commit and push to GitHub? [y/N]     │
        │ • Appends dated entry to CHANGELOG.md           │
        │ • Executes `git commit` &amp;amp; `git push`            │
        └─────────────────────────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Core Technical Stack &amp;amp; Engineering Decisions
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Local Inference Engine (&lt;code&gt;engine.py&lt;/code&gt; + Ollama + &lt;code&gt;qwen2.5:7b&lt;/code&gt;):&lt;/strong&gt;
Built a clean Python 3.14 wrapper around Ollama's local &lt;code&gt;/api/generate&lt;/code&gt; endpoint (&lt;code&gt;http://localhost:11434&lt;/code&gt;). By enforcing &lt;code&gt;"format": "json"&lt;/code&gt; with &lt;code&gt;temperature: 0.2&lt;/code&gt; and a tuned context window (&lt;code&gt;num_ctx: 2048&lt;/code&gt;), &lt;code&gt;qwen2.5:7b&lt;/code&gt; acts as a deterministic structured-data engine rather than a chatty chatbot, returning reliable JSON dictionaries parsed via &lt;code&gt;json.loads()&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Solving Local LLM Read Timeouts with Hybrid Delta Slicing (&lt;code&gt;Tidy.py&lt;/code&gt;):&lt;/strong&gt;
During initial testing, sending two full multi-page PDFs simultaneously to a 7B model on laptop hardware caused Ollama to exceed a 90-second HTTP &lt;code&gt;ReadTimeout&lt;/code&gt;. Worse, deleting the shorter file before inference finished risked losing data on timeout. I solved this with a 3-part fix:

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Fast CPU Filtering:&lt;/strong&gt; Used &lt;code&gt;PyMuPDF&lt;/code&gt; (&lt;code&gt;pymupdf&lt;/code&gt;) to extract text and Python's &lt;code&gt;difflib.SequenceMatcher&lt;/code&gt; on the first 1,500 characters to detect $\ge 55\%$ overlap in milliseconds without invoking the LLM.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Targeted Delta Slicing:&lt;/strong&gt; Instead of making &lt;code&gt;qwen2.5:7b&lt;/code&gt; read identical first pages twice, Python slices out the exact &lt;strong&gt;extra characters&lt;/strong&gt; present in the fuller PDF (&lt;code&gt;fuller["text"][shorter["chars"] : shorter["chars"] + 2000]&lt;/code&gt;) and sends only the overview + delta to Qwen. This made comparisons &lt;strong&gt;3x faster&lt;/strong&gt; and eliminated timeouts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Transactional File Deletion:&lt;/strong&gt; Moved &lt;code&gt;os.remove()&lt;/code&gt; strictly &lt;em&gt;after&lt;/em&gt; JSON inference succeeds and the user confirms &lt;code&gt;[y/N]&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Self-Healing Markdown Index (&lt;code&gt;SLOTH_INDEX.md&lt;/code&gt;):&lt;/strong&gt;
Instead of blindly appending text to a Markdown file (which leaves broken links when an inferior duplicate PDF gets replaced later), &lt;code&gt;Tidy.py&lt;/code&gt; stores structured document metadata in &lt;code&gt;sloth_memory.json&lt;/code&gt; and cleanly rebuilds &lt;code&gt;SLOTH_INDEX.md&lt;/code&gt; on every run.&lt;/li&gt;
&lt;/ol&gt;




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

&lt;p&gt;Building SlothSuite on open-weight models (&lt;code&gt;qwen2.5:7b&lt;/code&gt;), Ollama, and open-source Python libraries (&lt;code&gt;PyMuPDF&lt;/code&gt;, &lt;code&gt;watchdog&lt;/code&gt;) unlocked three capabilities that closed cloud APIs cannot offer:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Absolute Privacy for Personal Documents &amp;amp; Source Code:&lt;/strong&gt;
Real folders contain personal resumes (&lt;code&gt;main.pdf&lt;/code&gt;), university assignments, exam banks, and financial receipts, while Git repositories contain private codebases. Sending personal files and uncommitted &lt;code&gt;git diff&lt;/code&gt; logs to a commercial cloud LLM exposes sensitive data to third-party servers. Running &lt;code&gt;qwen2.5:7b&lt;/code&gt; on &lt;code&gt;localhost:11434&lt;/code&gt; guarantees &lt;strong&gt;0 bytes ever leave my machine&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero Per-Token Anxiety on Multi-Document Batches:&lt;/strong&gt;
Pre-scanning batches of PDFs, comparing versions, and analyzing Git diffs on every commit consumes thousands of tokens daily. With a metered API, you hesitate to run a batch script twice because every test run costs money. Open-weight local inference turns AI document deduplication and Git automation into a free utility you can run endlessly for &lt;strong&gt;$0.00&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;100% Offline Reliability:&lt;/strong&gt;
Whether working on shaky campus Wi-Fi or completely offline, SlothSuite categorizes PDFs, diffs duplicate versions, and commits code with zero internet dependency.&lt;/li&gt;
&lt;/ol&gt;




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

&lt;p&gt;I used &lt;strong&gt;SlothSuite&lt;/strong&gt; throughout this week both to clean up my own college study materials and to commit and push its own code for this challenge:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Deduplicating My Study Vault:&lt;/strong&gt; I dropped a batch of files into &lt;code&gt;./Downloaded_PDFs&lt;/code&gt;, including my Python OOP assignment (&lt;code&gt;Python_Assignment_2026 (1).pdf&lt;/code&gt;), a Data Scientist resume (&lt;code&gt;main.pdf&lt;/code&gt;), an LEEE 2026 question bank, and multiple overlapping DBMS notes (Normalization, Indexing Optimization, and Transactions &amp;amp; Concurrency). &lt;code&gt;Tidy.py&lt;/code&gt; automatically caught that &lt;code&gt;dbms_pdf4.pdf&lt;/code&gt; (1 page, 1,144 chars) was an incomplete duplicate of &lt;code&gt;dbms_pdf3.pdf&lt;/code&gt; (2 pages, 2,715 chars), highlighted the exact deadlock detection and timestamp ordering sections present in the fuller copy, deleted the inferior file on my confirmation, and extracted my &lt;strong&gt;February 28, 2026&lt;/strong&gt; OOP assignment deadline into &lt;code&gt;SLOTH_INDEX.md&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Letting the Sloth Commit Its Own Repo:&lt;/strong&gt; When refactoring the repository from &lt;code&gt;SlothInbox&lt;/code&gt; to &lt;code&gt;SlothVault&lt;/code&gt;, I didn't write a single commit message or changelog line manually. I ran &lt;code&gt;git add -A&lt;/code&gt; followed by &lt;code&gt;python sloth.py&lt;/code&gt;, and watched &lt;code&gt;qwen2.5:7b&lt;/code&gt; summarize 324 removed lines and 99 added lines into a clean &lt;code&gt;refactor(tidy): ...&lt;/code&gt; commit, update &lt;code&gt;CHANGELOG.md&lt;/code&gt;, and push itself to GitHub.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What's Next? — Turning SlothSuite into a CLI 🦥
&lt;/h2&gt;

&lt;p&gt;Right now, SlothSuite works through separate Python scripts. My next goal is to bring everything under one command-line interface, so the entire workflow feels like a single tool rather than a collection of scripts.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. &lt;code&gt;sloth --tidy&lt;/code&gt; — Smart Folder Cleanup
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;sloth &lt;span class="nt"&gt;--tidy&lt;/span&gt; &lt;span class="s2"&gt;"./Downloaded_PDFs"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I want to make document cleanup as simple as pointing Sloth at a folder. It would scan the PDFs, detect overlapping or incomplete duplicates, compare their contents, organize the surviving files, and generate or update the linked &lt;code&gt;SLOTH_INDEX.md&lt;/code&gt; dashboard.&lt;/p&gt;

&lt;p&gt;Before deleting, moving, or renaming anything, Sloth would present a full review and ask for explicit &lt;strong&gt;[y/N] confirmation&lt;/strong&gt;. No silent deletions — the user stays in control.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. &lt;code&gt;sloth git&lt;/code&gt; — A Self-Reviewing Git Assistant
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;sloth git
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The goal is to evolve the Git assistant beyond generating a commit message from staged changes. It would review the repository's existing commit history, inspect the current changes, and flag inconsistencies between the code, commit messages, and changelog.&lt;/p&gt;

&lt;p&gt;It could also suggest improvements to commit messages, identify opportunities to polish an existing &lt;code&gt;README.md&lt;/code&gt; or generate one if it's missing, and prepare the final changes for review before pushing to GitHub.&lt;/p&gt;

&lt;p&gt;Any history rewriting, commit amendments, or other potentially destructive Git operations would require explicit approval. The final push would also happen only after the user reviews and approves the proposed changes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The bigger vision:&lt;/strong&gt; one local-first CLI that keeps both your digital workspace and your Git repository organized — while letting a local AI handle the repetitive work and leaving the important decisions to you.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>opensource</category>
      <category>automation</category>
    </item>
    <item>
      <title>"VaultLedger: an offline invoice and bank statement auditor built on Ollama"</title>
      <dc:creator>Akshat Sharma</dc:creator>
      <pubDate>Sun, 18 Oct 2026 03:54:11 +0000</pubDate>
      <link>https://dev.to/akshat_slaya/vaultledger-an-offline-invoice-and-bank-statement-auditor-built-on-ollama-2kp7</link>
      <guid>https://dev.to/akshat_slaya/vaultledger-an-offline-invoice-and-bank-statement-auditor-built-on-ollama-2kp7</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;VaultLedger is an offline invoice and bank statement auditor for small shops and the bookkeepers who handle their records.&lt;/p&gt;

&lt;p&gt;I've been into a lot of local stores, and many of them already do their auditing on a PC, but none of it is automatic. Someone sits there matching invoices to bank statement lines one at a time. Internet isn't something you can count on either, so a cloud tool isn't an option. VaultLedger does the matching on the shop's own computer, with no internet needed.&lt;/p&gt;

&lt;p&gt;You upload the invoice PDFs and the bank statement CSV. A local open-weight model reads each invoice, and the app flags three things:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Duplicate invoices&lt;/strong&gt;, so the same bill isn't paid twice&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Invoices with no matching bank charge&lt;/strong&gt;, so you can see what hasn't been paid&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Bank charges with no invoice&lt;/strong&gt;, the missing receipts that usually show up at tax time&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;On my sample files the whole audit takes about a minute and a half to two minutes on my laptop, and the result downloads as an Excel report that can be handed to the owner or the client.&lt;/p&gt;

&lt;p&gt;Who would pay for this: a small shop owner or bookkeeper who spends hours matching invoices to bank statements by hand. One idea I'd want to test is a one-time license, or a small fee per batch of invoices. I haven't validated any price yet.&lt;/p&gt;

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

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/45cX5BlLzV0" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;The video runs the audit with my wifi turned off. It reads three invoices, flags the duplicate, and catches the one bank charge (QUICKPRINT STATIONERY, 86.50) that has no receipt.&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/akshat-bit420" rel="noopener noreferrer"&gt;
        akshat-bit420
      &lt;/a&gt; / &lt;a href="https://github.com/akshat-bit420/VaultLedger" rel="noopener noreferrer"&gt;
        VaultLedger
      &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;🔒 VaultLedger: 100% Offline PDF Invoice &amp;amp; Bank Auditor&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;VaultLedger&lt;/strong&gt; is a local-first AI financial reconciliation tool built for freelancers, independent consultants, and small agencies. It uses open-weight models via &lt;strong&gt;Ollama&lt;/strong&gt; (&lt;code&gt;qwen2.5:7b&lt;/code&gt; / &lt;code&gt;llama3.2:3b&lt;/code&gt;) to parse unstructured PDF invoices, cross-check them against bank statement CSVs, flag duplicate bills, and catch missing tax receipts—all running &lt;strong&gt;100% offline on &lt;code&gt;localhost&lt;/code&gt; with zero cloud APIs and zero data leaks&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;Built for the &lt;strong&gt;DEV Open-Source AI Challenge (Week 2: Side Hustle)&lt;/strong&gt;.&lt;/p&gt;

&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;✨ Key Features&lt;/h2&gt;
&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;100% Air-Gapped Privacy:&lt;/strong&gt; Sensitive financial PDFs, client names, tax IDs, and bank statements never leave your machine. Works completely without an internet connection.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic JSON Extraction:&lt;/strong&gt; Pairs &lt;strong&gt;Pydantic&lt;/strong&gt; with Ollama's native structured output API (&lt;code&gt;temperature=0.0&lt;/code&gt;) to extract &lt;code&gt;vendor_name&lt;/code&gt;, &lt;code&gt;invoice_number&lt;/code&gt;, &lt;code&gt;invoice_date&lt;/code&gt;, &lt;code&gt;subtotal&lt;/code&gt;, &lt;code&gt;tax_amount&lt;/code&gt;, &lt;code&gt;total_amount&lt;/code&gt;, &lt;code&gt;currency&lt;/code&gt;, and expense &lt;code&gt;category&lt;/code&gt; without hallucinations.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Duplicate Invoice Detection:&lt;/strong&gt; Automatically flags duplicate bills (same…&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/akshat-bit420/VaultLedger" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The repo includes a &lt;code&gt;sample_data&lt;/code&gt; folder with 3 fake invoices (one is a deliberate duplicate) and a bank CSV with one unreceipted charge, so you can test it in a couple of minutes.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Model:&lt;/strong&gt; &lt;code&gt;qwen2.5:7b&lt;/code&gt;, an open-weight model, running locally through Ollama&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;PDF reading:&lt;/strong&gt; PyMuPDF pulls the text out of each invoice&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Structured output:&lt;/strong&gt; a Pydantic schema is passed to Ollama's &lt;code&gt;format&lt;/code&gt; option, so the model has to return valid JSON with exactly the fields I need instead of free text I'd have to parse&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reconciliation:&lt;/strong&gt; pandas compares the extracted invoices to the bank CSV&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;UI:&lt;/strong&gt; Streamlit, with an Excel export of the audit&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The flow is simple on purpose: PDF text goes to the local model, the model returns structured JSON, and pandas does the matching and flagging.&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%2Fgozx2eeircgl150i5543.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%2Fgozx2eeircgl150i5543.png" alt="Diagram: invoice PDFs go through PyMuPDF to a local Ollama model, then pandas compares them with the bank CSV, and Streamlit shows the results and an Excel report" width="800" height="512"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Things that broke along the way, and what I learned:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;The category field was useless at first.&lt;/strong&gt; My schema asked for a category like Software or Travel, but I had typed the field as a plain string, so the model answered "Invoice" for everything. Ollama only enforces what is in the schema. Changing the field to a &lt;code&gt;Literal&lt;/code&gt; with the allowed values fixed it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The results vanished when I clicked download.&lt;/strong&gt; Streamlit reruns the whole script on every click, so the tables disappeared the moment I downloaded the Excel file. Storing the results in &lt;code&gt;st.session_state&lt;/code&gt; fixed it. I found it while recording my own demo, and I re-recorded the video after fixing it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;"Failed to connect to Ollama."&lt;/strong&gt; The first time I clicked the audit button, every invoice failed. The model was downloaded, but the Ollama server wasn't running yet. The fix is to start Ollama first (it runs in the background once installed). The app shows this error clearly, and the README tells the user to check that Ollama is running before starting the app. Running the AI locally means you also run the thing that serves it. That's the trade-off I accept in exchange for privacy and no per-document cost.&lt;/li&gt;
&lt;/ol&gt;

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

&lt;p&gt;&lt;strong&gt;Privacy is the product.&lt;/strong&gt; Shop owners and bookkeepers hold bank statements and invoices that other people trust them with. With an open-weight model running on the same machine, the documents never go to a server they don't control, and I could show it by recording the demo with wifi off.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It works without internet.&lt;/strong&gt; In the shops I've visited, you can't count on a connection. A cloud API simply stops working when the internet does. A local model doesn't care.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It costs nothing per invoice.&lt;/strong&gt; After the one-time model download, there is no API bill. A cloud API charges per document, which would eat into a small business's margin as volume grows.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I can swap models.&lt;/strong&gt; The model is a setting in the app, and any model Ollama can run can be dropped in. I only tested &lt;code&gt;qwen2.5:7b&lt;/code&gt; for this submission, so I can't claim how other models compare.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where a closed model would still win:&lt;/strong&gt; scanned or handwritten invoices. My version only reads PDFs that have a text layer, and a closed vision model would handle scans better today.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limitations (what I'd fix next)
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Matching is by amount only.&lt;/strong&gt; Two different vendors charging the same amount could falsely match. The next step is amount plus a date window plus fuzzy vendor-name matching.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Duplicates are detected by vendor, total and date.&lt;/strong&gt; Real duplicates often show up with the same invoice number, which I'd add next.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The model's numbers aren't verified.&lt;/strong&gt; A 7B model can get arithmetic wrong, so I'd add a check that subtotal plus tax equals total, and flag the row for review if it doesn't.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Scanned PDFs aren't supported yet.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;I tested on clean, synthetic invoices with one model.&lt;/strong&gt; I can't claim accuracy on messy real-world invoices yet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GST support&lt;/strong&gt; for Indian invoices is the next feature I'd build.&lt;/li&gt;
&lt;li&gt;It needs a computer that can run a 7B model comfortably.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This is a prototype built in a week. The part I trust is the idea: the documents stay on the machine, and the audit still works with no internet.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>opensource</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
