Skip to content
View abhigyanmahanta's full-sized avatar

Block or report abhigyanmahanta

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
abhigyanmahanta/README.md
Abhigyan Mahanta — I build the layer that refuses to trust the output

GitHub Website Instagram WhatsApp

RAAY Creative is open for work: automation, render pipelines, agent systems.
Two people. One builds the pipeline, one signs off the taste. I build the pipeline.


The problem I keep working on

Generative systems fail in a specific, nasty way: they produce output that looks right and is wrong.

A model repaints a product's packaging and the frame is still beautiful. A scraper reports a finished run and returns nothing. A deploy ships the HTML while the content-hashed CSS never lands, so the page is live and broken for an hour before anyone notices. In every case the naive check — does it look fine? did it say success? — passes.

So most of what I build is the layer that refuses to trust the output. The gate, the verifier, the ledger that records what was actually measured instead of what was hoped for.


Six repositories, one claim — each refuses to trust a different thing

🚀 Featured Work

Verifies that AI-generated product imagery actually preserved the real packaging. SIFT/RANSAC inlier gate, PASS/FAIL per pack per frame — because "looks fine" is not a check. Repairs the frame by homography when the scene is right and the print drifted.

Python OpenCV Tested

The studio site. Next.js on Cloudflare, ~11.5k LOC, 9/9 render tests — and a 540-line deploy verifier written after the live site served for an hour with every stylesheet 404ing behind a 200 response.

TypeScript Next.js Live

Frame-accurate video cutting from a written instruction. The model decides what should happen and never touches a timecode; perception owns every frame number, and resolution is pure lookup.

Python ffmpeg Whisper

Four research pipelines — Instagram, Pinterest, LinkedIn, Reels — sharing one discover → analyze → synthesize → render contract. 12,100 lines. Ships the scrapers and never the scraped corpus.

Python Playwright

An unattended B2B pipeline with staged execution, dedupe-by-inbox and a control plane that refuses to run on a failed day. Published with its own postmortem: 227 contacted, 0 replies tracked.

Python Postmortem

Append-only event log for agent pipelines. The dashboard infers which pipeline actually ran from the observed stage sequence. The logger never blocks, never fails a caller, and always exits 0.

Python Observability


🛠️ Toolkit

Build & automation Python TypeScript Node.js Bash Playwright

Media & vision OpenCV FFmpeg NumPy Whisper

AI & agents Claude Claude Code Gemini

Web & infra Next.js React Cloudflare Drizzle


📐 How I build

  • Instrumentation must never break the thing it instruments. The event logger never blocks, never fails a caller, and always exits 0 — so adding telemetry can't take down a render.
  • A number without a source is not a number. Every figure is tagged CONFIRMED, PROXY, CANNOT CONFIRM or TARGET. A figure that can't be sourced gets refused, not rounded.
  • Verify at the boundary the failure actually crosses. Not where it's convenient to assert.
  • Publish the refutation too. The research repos ship a findings file naming what was disproved, including when it contradicts the brief that commissioned it.

📫 Connect

Working alongside @anishadogra — she writes the rules, I build the systems that enforce them.

📍 New Delhi, India · raaycreative.com · Every graphic on this page is generated, not stock.

Pinned Loading

  1. agent-telemetry agent-telemetry Public

    Append-only event log for agent pipelines, plus a dashboard that infers which pipeline actually ran from the observed stage sequence.

    Python 2

  2. hub-pipelines hub-pipelines Public

    Four research pipelines — Instagram, Pinterest, LinkedIn, Reels — sharing one discover/analyze/synthesize/render contract. Scrapers only, no harvested corpus.

    Python 2

  3. lead-engine lead-engine Public

    Unattended B2B lead pipeline with staged execution, dedupe-by-inbox, and a control plane that refuses to run on a failed day. Published with its failure analysis.

    Python 2

  4. packfix packfix Public

    Verify AI-generated product imagery actually preserved the real packaging — SIFT/RANSAC inlier gate, PASS/FAIL per pack per frame.

    Python 2

  5. precision-video-editor precision-video-editor Public

    Frame-accurate video cutting from a written instruction — the model decides the edit and never touches a timecode.

    Python 2

  6. raaycreative.com raaycreative.com Public

    Source for raaycreative.com — Next.js on Cloudflare, with a deploy verifier written after a 200 lied and the site served an hour with every asset 404ing.

    TypeScript 2