<?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: Akul Bansal</title>
    <description>The latest articles on DEV Community by Akul Bansal (@akul_9512534b6f460d338911).</description>
    <link>https://dev.to/akul_9512534b6f460d338911</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%2F4198754%2F83891464-7d78-428b-9a07-96a86758f8e5.png</url>
      <title>DEV Community: Akul Bansal</title>
      <link>https://dev.to/akul_9512534b6f460d338911</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/akul_9512534b6f460d338911"/>
    <language>en</language>
    <item>
      <title>I Replaced $3,000 Eye-Trackers with a Laptop Webcam: Building a Hands-Free Spatial Communicator with SmartSpectra</title>
      <dc:creator>Akul Bansal</dc:creator>
      <pubDate>Sat, 14 Nov 2026 12:38:31 +0000</pubDate>
      <link>https://dev.to/akul_9512534b6f460d338911/i-replaced-3000-eye-trackers-with-a-laptop-webcam-building-a-hands-free-spatial-communicator-522l</link>
      <guid>https://dev.to/akul_9512534b6f460d338911/i-replaced-3000-eye-trackers-with-a-laptop-webcam-building-a-hands-free-spatial-communicator-522l</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;Traditional Augmentative and Alternative Communication (AAC) systems for people with severe motor and speech impairments (such as ALS, high-level quadriplegia, or locked-in syndrome) force an impossible trade-off: spend &lt;strong&gt;$3,000+ on specialized infrared eye-tracking hardware&lt;/strong&gt;, or endure exhausting &lt;strong&gt;dwell-time interfaces&lt;/strong&gt; where staring at a button for 1,000 ms just to register a single click leads to rapid visual fatigue and constant accidental selections (the classic &lt;em&gt;"Midas Touch"&lt;/em&gt; problem).&lt;/p&gt;

&lt;p&gt;Worse, existing gaze communicators suffer from two real-world bedside failures that nobody talks about:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Attention Theft:&lt;/strong&gt; The screen holds the user's eyes hostage. If they glance up to make eye contact with a loved one or try to vocalize, the cursor misfires and types gibberish.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One-Sided Visibility:&lt;/strong&gt; Caregivers and family members standing across the room can't read a laptop screen facing the patient.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;NeuroScribe Studio&lt;/strong&gt; turns any standard consumer webcam into a responsive, spatial assistive communication console by decoupling &lt;strong&gt;targeting&lt;/strong&gt; from &lt;strong&gt;activation&lt;/strong&gt;, introducing &lt;strong&gt;conversational standby gating&lt;/strong&gt;, and syncing messages in real time to a &lt;strong&gt;Caregiver Partner HUD (&lt;code&gt;/partner&lt;/code&gt;)&lt;/strong&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Core Capabilities
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;3D Spatial Gaze &amp;amp; Head-Pose Targeting:&lt;/strong&gt; High-contrast, multi-level hierarchical communication boards (&lt;em&gt;Physical Needs, Comfort &amp;amp; Position, People &amp;amp; Care, Quick Dialogue, Expressions, Emergency&lt;/em&gt;) driven by smoothed iris vectors and head pitch/yaw.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Temporal Blink Discrimination:&lt;/strong&gt; A client-side state machine separates natural involuntary blinks (&lt;code&gt;&amp;lt; 250 ms&lt;/code&gt;, ignored) from deliberate selection blinks (&lt;code&gt;350–850 ms&lt;/code&gt;, confirms card) and extended closures (&lt;code&gt;&amp;gt; 1,200 ms&lt;/code&gt;, undo/backspace).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Attentive Standby Lockout:&lt;/strong&gt; Automatically freezes card inputs the instant &lt;code&gt;face.talking&lt;/code&gt; is detected or the user tilts their head up (&lt;code&gt;pitch &amp;gt; 14°&lt;/code&gt;) to make natural eye contact with a caregiver.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Interactive 5-Second Calibration:&lt;/strong&gt; Dynamically normalizes iris bounding thresholds (&lt;code&gt;minX, maxX, minY, maxY&lt;/code&gt;) and blink timing windows to adapt to different seating angles, wheelchair heights, and eyelid control speeds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cross-Device Caregiver Mirror (&lt;code&gt;/partner&lt;/code&gt;):&lt;/strong&gt; Real-time WebSocket sync broadcasts composed phrases, spoken audio chimes, and haptic vibration alerts (&lt;code&gt;navigator.vibrate&lt;/code&gt;) to a caregiver's smartphone across the room—complete with two-way acknowledgment back to the patient.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Human-in-the-Loop Speech Synthesis:&lt;/strong&gt; Messages are never spoken automatically; users review their constructed sentence and confirm &lt;code&gt;[Speak Aloud]&lt;/code&gt; to trigger the browser's native Web Speech API.&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/uYGcQ-gCS0g" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;h3&gt;
  
  
  What Happens in the Demo:
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Consent &amp;amp; Privacy Gate:&lt;/strong&gt; Explicit pre-capture authorization explaining local in-memory processing and zero telemetry.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hands-Free Spatial Navigation:&lt;/strong&gt; Navigating the 3D spatial wave-cards via eye gaze and selecting phrases with deliberate blinks—with hands completely off the keyboard.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Attentive Standby Lockout:&lt;/strong&gt; Looking up and speaking freezes the grid so natural eye contact with a caregiver never causes accidental selections.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Caregiver Mirror Sync (&lt;code&gt;/partner&lt;/code&gt;):&lt;/strong&gt; Confirming &lt;strong&gt;Speak Aloud&lt;/strong&gt; broadcasts the message in real time to a second device with audio and haptic feedback.&lt;/li&gt;
&lt;/ol&gt;




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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Public GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/Neuroscribe" rel="noopener noreferrer"&gt;https://github.com/Akulbanxal/Neuroscribe&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;Rather than treating SmartSpectra as a passive telemetry display, &lt;strong&gt;NeuroScribe Studio&lt;/strong&gt; orchestrates four live Presage human signals in real time to power a hands-free, zero-fatigue assistive communication loop:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;Iris Vectors &amp;amp; 478 Facial Landmarks (&lt;code&gt;faceMetrics&lt;/code&gt;) → Spatial Card Targeting &amp;amp; 3D Parallax:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Normalized iris coordinates (&lt;code&gt;[0, 1]&lt;/code&gt;) and head pose (&lt;code&gt;pitch&lt;/code&gt; and &lt;code&gt;yaw&lt;/code&gt;) are streamed from the local SmartSpectra Node SDK daemon (&lt;code&gt;@smartspectra/node-sdk&lt;/code&gt;) at 30 FPS over WebSockets (&lt;code&gt;ws://localhost:8080&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;To eliminate raw webcam jitter, the frontend applies an &lt;strong&gt;Exponential Moving Average (EMA)&lt;/strong&gt; filter (α = 0.28–0.35) and maps the smoothed gaze vector onto calibrated 3×2 spatial wave-cards with deadzones, dynamically tilting the 3D glass deck in response to head orientation.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;Binary Blink Detection (&lt;code&gt;face.blinking&lt;/code&gt;) → Temporal Selection &amp;amp; Undo State Machine:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Instead of exhausting dwell-time timers (staring at a button for 1,000+ ms), NeuroScribe measures the exact duration of continuous &lt;code&gt;face.blinking.detected&lt;/code&gt; windows on the client side:

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Natural involuntary blinks (&lt;code&gt;&amp;lt; 250 ms&lt;/code&gt;):&lt;/strong&gt; Ignored completely to prevent accidental clicks.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deliberate selection blinks (&lt;code&gt;350–850 ms&lt;/code&gt;):&lt;/strong&gt; Fills a live neon progress bar on the targeted card and appends the phrase to the sentence builder.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Extended eye closure (&lt;code&gt;&amp;gt; 1,200 ms&lt;/code&gt;):&lt;/strong&gt; Triggers an instant &lt;strong&gt;Undo / Backspace&lt;/strong&gt; action.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;Speech Articulation Detection (&lt;code&gt;face.talking&lt;/code&gt;) &amp;amp; Head Pitch → Attentive Standby Lockout:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The moment &lt;code&gt;face.talking.detected&lt;/code&gt; flips to &lt;code&gt;true&lt;/code&gt;—or when the user lifts their chin (&lt;code&gt;pitch &amp;gt; 14°&lt;/code&gt;) to make natural eye contact with a bedside caregiver—the UI immediately enters &lt;strong&gt;Attentive Standby&lt;/strong&gt;, freezing all card selections and dimming the grid so real-world conversation never triggers false inputs.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;Respiratory Waveform (&lt;code&gt;breathing.upperTrace&lt;/code&gt;) &amp;amp; Facial Expressions (&lt;code&gt;faceMetrics.expressions&lt;/code&gt;) → Live HUD &amp;amp; Ambient Feedback:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Real-time upper-chest respiratory excursions drive the rhythmic breathing sparkline in the system status bar, while dominant facial expression probabilities subtly adapt the ambient spatial lighting.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  SDK Integration Architecture
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Backend Bridge (&lt;code&gt;server.mjs&lt;/code&gt;):&lt;/strong&gt; Initializes &lt;code&gt;SmartSpectraSDK&lt;/code&gt; locally via C++ FFI with &lt;code&gt;requestedMetrics: [...faceMetrics, ...breathingMetrics]&lt;/code&gt; and &lt;code&gt;enableTelemetry: false&lt;/code&gt;, normalizing all frame timestamps to epoch microseconds (&lt;code&gt;Date.now() * 1000&lt;/code&gt;) to guarantee continuous, stall-free streaming.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Frontend Consumer (&lt;code&gt;useBiometrics.ts&lt;/code&gt;):&lt;/strong&gt; Ingests the 30 FPS WebSocket JSON stream, applies EMA coordinate smoothing, runs the temporal blink discriminator, and synchronizes confirmed phrases to both the Web Speech API (&lt;code&gt;speechSynthesis&lt;/code&gt;) and the cross-device &lt;strong&gt;Caregiver Mirror (&lt;code&gt;/partner&lt;/code&gt;)&lt;/strong&gt;.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;p&gt;I am entering &lt;strong&gt;NeuroScribe Studio&lt;/strong&gt; for &lt;strong&gt;Best Thoughtful Real-World Use&lt;/strong&gt; (Primary) as well as &lt;strong&gt;Best Overall Human-Sensing Experience&lt;/strong&gt;.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The Real-World Need:&lt;/strong&gt; Millions of individuals living with severe motor and speech impairments—such as ALS, high-level spinal cord injuries, cerebral palsy, or brainstem stroke—lose the ability to speak or use touchscreens. Existing Augmentative and Alternative Communication (AAC) setups either cost &lt;strong&gt;$3,000+ for dedicated infrared hardware&lt;/strong&gt; or rely on exhausting &lt;strong&gt;dwell-time timers&lt;/strong&gt; (staring at a button for 1,000+ ms to click) that cause rapid ocular fatigue and constant accidental inputs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Who It’s For:&lt;/strong&gt; Non-speaking or motor-impaired individuals who need a reliable, zero-cost way to express urgent needs, comfort adjustments, and everyday thoughts using only a standard laptop webcam—alongside their &lt;strong&gt;bedside caregivers and family members&lt;/strong&gt; who need immediate visibility across the room.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;How It Was Designed for Real-World Usability:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Zero-Fatigue Temporal Blink Activation:&lt;/strong&gt; Decouples eye-gaze targeting from intentional blink confirmation (&lt;code&gt;350–850 ms&lt;/code&gt;), ignoring involuntary natural blinks (&lt;code&gt;&amp;lt; 250 ms&lt;/code&gt;) and mapping extended closures (&lt;code&gt;&amp;gt; 1,200 ms&lt;/code&gt;) to &lt;strong&gt;Undo&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Attentive Standby Lockout:&lt;/strong&gt; Automatically freezes card inputs when &lt;code&gt;face.talking&lt;/code&gt; is detected or when the user tilts their head up (&lt;code&gt;pitch &amp;gt; 14°&lt;/code&gt;) to make natural eye contact with a loved one—eliminating accidental selections during real-world conversation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;5-Second Adaptive Calibration &amp;amp; High-Contrast Spatial Cards:&lt;/strong&gt; Large, unmistakable 3×2 wave-cards with customizable blink thresholds and seating calibration accommodate varying wheelchair angles and eyelid control speeds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cross-Device Caregiver Mirror (&lt;code&gt;/partner&lt;/code&gt;):&lt;/strong&gt; Broadcasts composed phrases, spoken audio chimes, and haptic vibrations (&lt;code&gt;navigator.vibrate&lt;/code&gt;) to a caregiver's smartphone in real time, complete with two-way acknowledgment back to the patient.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ul&gt;




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

&lt;p&gt;&lt;strong&gt;NeuroScribe Studio&lt;/strong&gt; was architected from the ground up with strict privacy-first guardrails and full compliance with the challenge rules:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Explicit Pre-Capture Consent:&lt;/strong&gt; Camera access and WebSocket biometric streaming remain completely disabled at launch until the user explicitly reviews and accepts the on-screen &lt;strong&gt;Camera Consent &amp;amp; Privacy Notice Modal&lt;/strong&gt; (with a one-click option to revoke/pause capture at any time).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;100% Local In-Memory Processing (Zero Storage / Zero Sharing):&lt;/strong&gt; Video frames are captured locally by the Node.js daemon (&lt;code&gt;@smartspectra/node-sdk&lt;/code&gt;), processed strictly in volatile RAM, and immediately discarded. &lt;strong&gt;Zero video frames, facial images, landmark recordings, or biometric logs are ever saved to disk, persisted in databases, or transmitted to external servers.&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Retention &amp;amp; Deletion:&lt;/strong&gt; Because all processing occurs ephemerally in volatile memory with zero disk persistence, data retention is &lt;strong&gt;0 seconds&lt;/strong&gt;—nothing is stored, so no residual user data ever remains after closing the session.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;SDK Telemetry Disabled:&lt;/strong&gt; The SmartSpectra SDK instance is explicitly initialized with &lt;code&gt;enableTelemetry: false&lt;/code&gt;, ensuring no diagnostic session telemetry or usage metrics leave the local machine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No Third-Party Cloud AI Services:&lt;/strong&gt; All gaze smoothing, blink classification, sentence assembly, and text-to-speech synthesis (&lt;code&gt;window.speechSynthesis&lt;/code&gt;) run 100% locally on the user's device without calling external AI APIs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Non-Medical &amp;amp; Non-Diagnostic Disclosure:&lt;/strong&gt; NeuroScribe Studio is strictly an &lt;strong&gt;assistive human-computer interaction (HCI) communication interface&lt;/strong&gt;. It makes &lt;strong&gt;no medical, diagnostic, or treatment claims&lt;/strong&gt;, performs no clinical decision-making, and does not enable covert or non-consensual monitoring.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Agent Session
&lt;/h2&gt;

&lt;p&gt;&lt;em&gt;Built solo using React 19, TypeScript, Vite, Tailwind CSS, and &lt;code&gt;@smartspectra/node-sdk&lt;/code&gt;, with AI-assisted architectural iteration and UI prototyping.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>presagechallenge</category>
      <category>webdev</category>
      <category>showdev</category>
    </item>
    <item>
      <title>RepoGraveyard 3D: Resurrecting Legacy Code with Gemma 4</title>
      <dc:creator>Akul Bansal</dc:creator>
      <pubDate>Sat, 31 Oct 2026 18:24:16 +0000</pubDate>
      <link>https://dev.to/akul_9512534b6f460d338911/repograveyard-3d-resurrecting-legacy-code-with-gemma-4-412g</link>
      <guid>https://dev.to/akul_9512534b6f460d338911/repograveyard-3d-resurrecting-legacy-code-with-gemma-4-412g</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;RepoGraveyard 3D is an atmospheric software archaeology workstation built for the "Ghost in the Machine" challenge. Every forgotten open-source repository carries a ghost—abandoned TODOs, breaking dependency changes, and bit-rotted toolchains. &lt;/p&gt;

&lt;p&gt;RepoGraveyard 3D identifies and repairs legacy open-source bit-rot using local &lt;strong&gt;Gemma 4 9B&lt;/strong&gt; AI orchestrated inside an interactive retro 3D WebGL CRT server room.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Live Interactive Workstation&lt;/strong&gt;: &lt;a href="https://banxal.tech" rel="noopener noreferrer"&gt;https://banxal.tech&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/rYTAvjV03Y4" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;GitHub Repository&lt;/strong&gt;: &lt;a href="https://github.com/Akulbanxal/graveyard-3D" rel="noopener noreferrer"&gt;https://github.com/Akulbanxal/graveyard-3D&lt;/a&gt;&lt;/p&gt;

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

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;3D WebGL Terminal: Built with React Three Fiber, Three.js, and custom post-processing GLSL glitch shaders.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Local AI Inference: Local Gemma 4 9B model (via Ollama) executing stack-trace diagnostics, toolchain failure analysis, and unified git diff generation.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Agent Orchestration Framework: Managed using the Mastra Agent Framework.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Observability &amp;amp; Tracing: Real-time agent tracing, token counts, and span latencies instrumented via @sentry/nextjs.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Infrastructure: Production Web Service deployed on Render bound to our custom .tech domain (banxal.tech).&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Narration: ElevenLabs synthesized ghost voiceover for screen walkthrough audio.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;

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

&lt;p&gt;Open-weight models like Gemma 4 make private, local software archaeology possible. Legacy codebases and commit histories often contain sensitive security logs, internal structural data, or personal details that developers cannot share with proprietary cloud APIs. Running Gemma 4 locally via Ollama guarantees data sovereignty, zero API costs, and full offline resilience while granting full control over system prompts and agent execution parameters.&lt;/p&gt;

&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;This application was designed and iterated alongside agentic AI coding assistants. Below is an authentic execution trace payload generated during local Gemma 4 code exhumation: {&lt;br&gt;
  "target_repo": "octocat/hello-world",&lt;br&gt;
  "status": "RESURRECTED",&lt;br&gt;
  "agent_framework": "Mastra v1.0",&lt;br&gt;
  "model": "gemma-4-9b-instruct",&lt;br&gt;
  "telemetry": {&lt;br&gt;
    "sentry_trace_id": "8f3b21a09c4d4e12",&lt;br&gt;
    "execution_latency_ms": 1420&lt;br&gt;
  },&lt;br&gt;
  "diagnostics": {&lt;br&gt;
    "bit_rot_cause": "Deprecated peer dependency resolution manifest",&lt;br&gt;
    "patch_status": "Clean git diff synthesized"&lt;br&gt;
  }&lt;br&gt;
}&lt;/p&gt;

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

&lt;p&gt;Software archaeology is about remembering the human story behind the code. Stepping into the 3D server room with flickering CRT monitors, low ectoplasm fog, and the low hum of local AI inference transforms routine debugging into an atmospheric experience—giving forgotten open-source projects a second life.&lt;/p&gt;

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

&lt;ol&gt;
&lt;li&gt;Overall Winner: "Ghost in the Machine" interactive 3D WebGL workstation implementation.&lt;/li&gt;
&lt;li&gt;Best Use of Gemma: Local Gemma 4 9B model handling automated code repair and diagnosis.&lt;/li&gt;
&lt;li&gt;Best Use of Render: Production app runtime and Next.js frontend deployed live on Render.&lt;/li&gt;
&lt;li&gt;Best Use of Mastra: Workflow orchestration across agent execution steps.&lt;/li&gt;
&lt;li&gt;Best Use of Sentry Agent Tracing: Distributed tracing of model latency and agent spans.&lt;/li&gt;
&lt;li&gt;Best Use of a .tech Domain: Hosted live at banxal.tech.&lt;/li&gt;
&lt;li&gt;Best Use of GitHub Copilot: Built with Copilot CLI and GitHub Actions workflows.&lt;/li&gt;
&lt;li&gt;Best Use of ElevenLabs: Demo video voiceover narration.&lt;/li&gt;
&lt;li&gt;Best Use of Entire: Linked agent prompt session traces.&lt;/li&gt;
&lt;/ol&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>mastra</category>
    </item>
    <item>
      <title>GREENWASH: Catching AI Coding Agents When Tests Lie</title>
      <dc:creator>Akul Bansal</dc:creator>
      <pubDate>Sun, 25 Oct 2026 13:41:44 +0000</pubDate>
      <link>https://dev.to/akul_9512534b6f460d338911/greenwash-catching-ai-coding-agents-when-tests-lie-3hed</link>
      <guid>https://dev.to/akul_9512534b6f460d338911/greenwash-catching-ai-coding-agents-when-tests-lie-3hed</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;A green checkmark usually means good news. Your tests pass, your build succeeds, and you can finally move on to the next task.&lt;/p&gt;

&lt;p&gt;But what if the bug was never actually fixed?&lt;/p&gt;

&lt;p&gt;Imagine asking an AI coding agent to fix a payment bug. Instead of correcting the payment logic, it weakens the test assertion so incorrect results are accepted. The test passes, but the underlying problem might still be there.&lt;/p&gt;

&lt;p&gt;That is the problem I wanted to explore with &lt;strong&gt;GREENWASH&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;GREENWASH is a code-integrity auditing project designed to identify suspicious changes in AI-generated code, particularly changes that make software appear correct by weakening the checks used to verify it.&lt;/p&gt;

&lt;p&gt;The project focuses on six scenarios:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Weakened assertions:&lt;/strong&gt; A test becomes less strict and starts accepting incorrect behaviour.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deleted failing tests:&lt;/strong&gt; A test exposing a defect disappears instead of the defect being fixed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Skipped tests:&lt;/strong&gt; An important test is disabled rather than repaired.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Excessive mocking:&lt;/strong&gt; Fake dependencies replace meaningful checks of real functionality.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Weakened CI quality gates:&lt;/strong&gt; Required checks are removed or bypassed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Honest bug fixes:&lt;/strong&gt; The implementation is corrected while preserving the intended requirements.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The idea is simple: don't just ask whether the tests passed. Ask what changed to make them pass.&lt;/p&gt;

&lt;p&gt;GREENWASH aims to compare code and test changes, identify suspicious patterns, and use relevant behavioural checks to evaluate whether the original requirements are still being verified.&lt;/p&gt;

&lt;p&gt;It is not about assuming that AI agents deliberately cheat. It is about making potentially risky changes visible and giving developers evidence they can investigate.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;Live demo:&lt;/strong&gt; [&lt;a href="https://banxal.tech/#landing" rel="noopener noreferrer"&gt;https://banxal.tech/#landing&lt;/a&gt;]&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Demo video:&lt;/strong&gt; [&lt;a href="https://youtu.be/Etl5xBcXUF0" rel="noopener noreferrer"&gt;https://youtu.be/Etl5xBcXUF0&lt;/a&gt;]&lt;/p&gt;

&lt;p&gt;The most useful demonstration is to compare two versions of the same codebase.&lt;/p&gt;

&lt;p&gt;In the first version, a test catches a genuine defect. In the modified version, the test is weakened or removed instead of fixing the underlying problem.&lt;/p&gt;

&lt;p&gt;GREENWASH should then highlight the suspicious change and show the relevant verification evidence supported by the implementation.&lt;/p&gt;

&lt;p&gt;A second scenario demonstrates the opposite: a genuine bug fix that preserves the intended behaviour and passes the configured checks.&lt;/p&gt;

&lt;p&gt;Together, these scenarios show why a passing test suite should be examined in context, rather than treated as automatic proof of correctness.&lt;/p&gt;

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

&lt;p&gt;&lt;strong&gt;GitHub repository:&lt;/strong&gt; [&lt;a href="https://github.com/Akulbanxal/GreenWash" rel="noopener noreferrer"&gt;https://github.com/Akulbanxal/GreenWash&lt;/a&gt;]&lt;/p&gt;

&lt;p&gt;The repository contains the implementation, detection logic, and supporting project files. It is the best place to explore the code and see which of the planned audit behaviours are currently implemented.&lt;/p&gt;

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

&lt;p&gt;I approached GREENWASH as a problem of trust and verification rather than simply another testing dashboard.&lt;/p&gt;

&lt;p&gt;The intended audit workflow has four main stages:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Inspect the changes.&lt;/strong&gt; Compare a trusted baseline with the modified code.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Identify suspicious patterns.&lt;/strong&gt; Examine test files and CI configuration for changes such as weakened assertions, removed tests, skip directives, excessive mocking, and weakened quality gates.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Verify relevant behaviour.&lt;/strong&gt; Where independent tests or behavioural checks are available, use them to evaluate whether the original requirements are still satisfied.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Present the evidence.&lt;/strong&gt; Explain which changes triggered findings and show the verification results that support them.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;A key design challenge is distinguishing suspicious changes from legitimate ones. A deleted test might be redundant, and mocking can be appropriate in many situations. A useful auditor therefore needs context rather than a simplistic rule that labels every test modification as bad.&lt;/p&gt;

&lt;p&gt;The project also includes a visual interface for exploring audit scenarios and their findings. I have been focusing on the graphics integration and overall presentation so that the underlying concepts are easier to understand without losing sight of the technical evidence.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Technology used:&lt;/strong&gt; [Add the actual languages, frameworks, libraries, and tools used in your implementation.]&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Open-source AI used:&lt;/strong&gt; [Name the actual open-source model, agent framework, or AI tooling used, and explain exactly how it contributes to GREENWASH.]&lt;/p&gt;

&lt;p&gt;I want to be clear about an important limitation: passing the configured checks cannot prove that software is completely free of bugs. The goal is to improve the quality of the review and make suspicious verification changes easier to investigate.&lt;/p&gt;

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

&lt;p&gt;AI-assisted development is moving quickly, but developers still need ways to understand and verify the code these tools produce.&lt;/p&gt;

&lt;p&gt;Open innovation makes it possible for more developers to inspect implementations, question design decisions, experiment with different models, and adapt tools to their own workflows.&lt;/p&gt;

&lt;p&gt;For a project like GREENWASH, this matters because trust cannot depend entirely on a tool claiming that everything is fine. Developers should be able to inspect the rules, understand the evidence, and challenge a finding when the context suggests a legitimate change.&lt;/p&gt;

&lt;p&gt;Open-source models and frameworks can also give developers more flexibility over how AI capabilities are integrated and deployed, depending on the chosen model, licence, infrastructure, and hardware.&lt;/p&gt;

&lt;p&gt;The broader goal is not to discourage the use of AI coding agents. It is to encourage a development workflow where their output can be examined critically.&lt;/p&gt;

&lt;p&gt;AI can help us write code faster. Better verification can help us understand whether that code deserves our confidence.&lt;/p&gt;

&lt;h2&gt;
  
  
  My Agent Session
&lt;/h2&gt;

&lt;p&gt;[Optional: Add your DevRelay agent session link or embed your recorded session using the challenge's supported &lt;code&gt;agent_session&lt;/code&gt; format.]&lt;/p&gt;

&lt;p&gt;A session showing how you explored the problem, developed the implementation, tested a scenario, and refined the result can help readers understand the decisions behind the project.&lt;/p&gt;

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

&lt;p&gt;The potential payoff is straightforward: less time spent manually searching through every test and configuration change, and more time investigating changes that genuinely deserve attention.&lt;/p&gt;

&lt;p&gt;GREENWASH aims to make the first pass of code-integrity review more systematic by highlighting suspicious modifications and presenting their supporting evidence.&lt;/p&gt;

&lt;p&gt;I don't have a measured time-saving figure to claim yet, and the actual benefit will depend on the repository, the rules, and the quality of the verification checks.&lt;/p&gt;

&lt;p&gt;But the motivation is practical: when AI can produce code quickly, reviewing what changed—and whether the verification remains meaningful—becomes increasingly important.&lt;/p&gt;

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

&lt;p&gt;GREENWASH is designed to make AI-generated code changes easier to audit, with a focus on transparent verification, flexible deployment, and traceable agent workflows. We are entering the following partner categories where our implementation and demonstrations meet the requirements.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Best Use of Render
&lt;/h3&gt;

&lt;p&gt;GREENWASH is deployed on Render, with a Docker-based service running the Express REST API and React frontend through a unified application deployment.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Live application:&lt;/strong&gt; &lt;a href="https://banxal.tech" rel="noopener noreferrer"&gt;https://banxal.tech&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Render service:&lt;/strong&gt; &lt;code&gt;greenwash.onrender.com&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deployment configuration:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/render.yaml" rel="noopener noreferrer"&gt;render.yaml&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Container configuration:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/Dockerfile" rel="noopener noreferrer"&gt;Dockerfile&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The deployment makes GREENWASH accessible through a public interface while keeping the application and audit workflow in one deployment.&lt;/p&gt;

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

&lt;p&gt;GREENWASH includes a semantic-review component designed to use Google's open-weight Gemma models to complement deterministic code-integrity checks.&lt;/p&gt;

&lt;p&gt;The reviewer analyses code-change context, explains suspicious modifications, and suggests additional edge cases for investigation. GREENWASH also supports a self-hosted configuration using Ollama, allowing a compatible Gemma model to run locally when the local inference service is configured.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Implementation:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/packages/core/src/gemma/GemmaSemanticReviewer.ts" rel="noopener noreferrer"&gt;GemmaSemanticReviewer.ts&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local inference configuration:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/docker-compose.yml" rel="noopener noreferrer"&gt;docker-compose.yml&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The goal is to combine deterministic checks with model-assisted analysis, while keeping the evidence behind an audit finding visible to the developer.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Note: The current public Render configuration uses a mock Gemma endpoint. Demonstrate a working local Ollama/Gemma run separately before claiming live model inference as part of the deployed service.&lt;/em&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Best Use of a .tech Domain
&lt;/h3&gt;

&lt;p&gt;GREENWASH is available through a custom &lt;code&gt;.tech&lt;/code&gt; domain:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Live URL:&lt;/strong&gt; &lt;a href="https://banxal.tech" rel="noopener noreferrer"&gt;https://banxal.tech&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Using a custom domain gives the project a stable public entry point for demonstrating its interface and audit scenarios.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Best Use of Mastra
&lt;/h3&gt;

&lt;p&gt;GREENWASH uses Mastra's typed workflow engine to organise its audit pipeline into distinct stages.&lt;/p&gt;

&lt;p&gt;The workflow coordinates scenario-diff validation, deterministic scanning, contract replay, semantic analysis, report normalisation, and trace emission.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Implementation:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/packages/core/src/mastra/workflow.ts" rel="noopener noreferrer"&gt;workflow.ts&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Breaking the audit into stages makes the workflow easier to inspect, test, and extend as additional detection rules and verification checks are introduced.&lt;/p&gt;

&lt;h3&gt;
  
  
  5. Best Use of Sentry Agent Tracing
&lt;/h3&gt;

&lt;p&gt;GREENWASH includes a Sentry integration for tracing audit-workflow execution and individual stages.&lt;/p&gt;

&lt;p&gt;The tracing code records workflow metadata, step information, result status, and errors. It also filters sensitive attributes and redacts secrets before telemetry is sent.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Implementation:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/packages/core/src/sentry/tracer.ts" rel="noopener noreferrer"&gt;Sentry tracer&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;This helps make failures easier to investigate and provides a way to understand how an audit progresses through the pipeline.&lt;/p&gt;

&lt;h3&gt;
  
  
  6. Best Use of MongoDB Atlas
&lt;/h3&gt;

&lt;p&gt;GREENWASH includes a MongoDB-backed persistence layer designed to store audit history and related records.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Implementation:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/packages/core/src/persistence/repository.ts" rel="noopener noreferrer"&gt;repository.ts&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The repository supports configuring a MongoDB connection through environment variables, allowing a suitable MongoDB Atlas deployment to be used for persistent audit storage.&lt;/p&gt;

&lt;p&gt;The important distinction is between having MongoDB integration in the code and demonstrating that a live Atlas cluster is actually connected and storing records. The latter should be shown for this prize category.&lt;/p&gt;

&lt;h3&gt;
  
  
  7. Best Use of Entire
&lt;/h3&gt;

&lt;p&gt;GREENWASH includes an agent-session provider that can parse session records and extract relevant provenance information for investigation.&lt;/p&gt;

&lt;p&gt;The implementation also supports filtering session content and redacting secrets, helping the audit workflow use agent history without unnecessarily exposing sensitive information.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Implementation:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/packages/core/src/provenance/EntireSessionProvider.ts" rel="noopener noreferrer"&gt;EntireSessionProvider.ts&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Agent provenance can provide useful context when investigating suspicious sequences of development actions, such as changes made around a failing test.&lt;/p&gt;

&lt;p&gt;For this category, we will demonstrate the session ingestion flow with a real, supported Entire session rather than relying only on the presence of a parser in the repository.&lt;/p&gt;

&lt;h3&gt;
  
  
  8. Best Use of GitHub Copilot
&lt;/h3&gt;

&lt;p&gt;GREENWASH includes repository-level development instructions and a GitHub Actions workflow that runs integrity-audit checks as part of the development process.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Copilot instructions:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/.github/copilot-instructions.md" rel="noopener noreferrer"&gt;.github/copilot-instructions.md&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CI workflow:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/.github/workflows/ci.yml" rel="noopener noreferrer"&gt;.github/workflows/ci.yml&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The intention is to make integrity checks part of the pull-request workflow, helping identify suspicious test or verification changes before they are merged.&lt;/p&gt;

&lt;p&gt;We will demonstrate the actual Copilot or GitHub Actions workflow used for this submission to show how the integration contributes to the project.&lt;/p&gt;

&lt;h3&gt;
  
  
  9.Additional Deployment Target: DigitalOcean
&lt;/h3&gt;

&lt;p&gt;GREENWASH also includes a DigitalOcean App Platform specification and Docker Compose configuration to support deployment in a self-managed environment.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;App Platform configuration:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/.do/app.yaml" rel="noopener noreferrer"&gt;.do/app.yaml&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Docker Compose configuration:&lt;/strong&gt; &lt;a href="https://github.com/Akulbanxal/GreenWash/blob/main/docker-compose.yml" rel="noopener noreferrer"&gt;docker-compose.yml&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These files provide a starting point for deployment and local-service orchestration. We will list &lt;strong&gt;Best Use of DigitalOcean&lt;/strong&gt; as a prize entry only after demonstrating a working deployment or another qualifying use of DigitalOcean's services.&lt;/p&gt;




&lt;p&gt;Our aim is to make GREENWASH more than a dashboard: it is an evidence-first approach to examining AI-generated code changes, with inspectable workflows, configurable integrations, and verification results that developers can investigate.&lt;/p&gt;

&lt;p&gt;GREENWASH is my exploration of a question I think every team using AI-assisted development should keep asking:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Did we fix the problem, or did we only make the problem harder to see?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I'd love to hear how others are approaching code review and verification in AI-assisted workflows.&lt;/p&gt;

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