Skip to content
View itsoumya-d's full-sized avatar
6 followers · 3 following

Achievements

Achievement: Pull Sharkx3Achievement: YOLOAchievement: QuickdrawAchievement: Pair Extraordinaire

Achievements

Achievement: Pull Sharkx3Achievement: YOLOAchievement: QuickdrawAchievement: Pair Extraordinaire

Block or report itsoumya-d

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. 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
itsoumya-d/README.md

Soumya Debnath, AI Product Engineer: live profile header

Email me LinkedIn Full engineering log

I build applied AI systems end to end: agent runtimes with safety rails, LLM gateways and evals, research-grade RL experiments, and consumer apps that are live on the App Store and Google Play. Everything here has tests, CI and an honest list of what is not claimed.


01 Live telemetry

Live activity, language mix and CI health


02 Hiring for a specific role? Pick a lens

🧠  AI / LLM platform  Gateways, routing and retrieval that make LLM features cheaper, observable and auditable.

LLM Control Plane: One gateway in front of every LLM provider, with failover, rate limits, cost tracking and an audit log. RouteLLM Gateway: Sends easy prompts to a cheap model and hard ones to a frontier model, so you pay less for the same answers.

GraphRAG CodeAudit: Turns a codebase into a knowledge graph so an AI can answer big-picture audit questions, not just keyword lookups.

🛡️  Agent safety & security  Rails for AI agents that touch real systems: approvals, policy checks and cryptographic receipts.

HEROS: A safety layer for AI agents: risky actions need a human OK, and every action leaves a tamper-proof receipt. HEROS Firewall: A security guard between an AI agent and its tools that blocks dangerous calls before they run.

HEROS Cloud: A workflow studio where the AI writes plain-English safety rules from its own failures, and a human approves them.

🔬  ML / RL research  Small, reproducible experiments that isolate one idea and measure it, with GPU-scale runs clearly marked as pending.

HEROS-RL: Trains a language model to rewrite risky database migrations into safe, zero-downtime ones. Reflexion-SWE: A coding agent that learns from its own failed attempts, without retraining the model.

RLVR Bias Probe: An experiment showing why a popular RL training method quietly rewards longer answers, and how a fix removes it.

⚡  Performance & edge  Making AI run fast and private: SIMD kernels, in-browser inference and post-quantum P2P.

PerfBench: Hand-tuned CPU kernels for the maths inside transformer models, each checked against a reference. EdgeInfer: Runs AI models and semantic search entirely in the browser, with no server.

PeerVault: Peer-to-peer file transfer in the browser with post-quantum encryption and no server in the middle.

📱  Mobile & product  Shipped consumer apps with real store listings, plus tooling that turns design references into UI.

CTrackAI: AI calorie tracker: snap a photo of your meal, scan a barcode, and track macros, fasting and water. Preeo: Cycle and health companion: log cycles and symptoms, see BBT charts, and share reports with a doctor.

Mobbin UI Compiler: Generates accessible React screens from real-app design references.

🧱  Backend, payments & integration  Correctness-first services: idempotency, ledgers, outbox patterns and industrial integrations.

go-ledger: A payment service that cannot double-charge: idempotent requests and a double-entry ledger that always balances. Factory Monitor: Software that watches factory machines and cameras in real time and plugs into ERP/MES systems.

AutoPilot FDE: Finds repetitive work in a company's emails and tickets, then drafts automations for it (prototype).


03 Flagship systems

HEROS: A safety layer for AI agents: risky actions need a human OK, and every action leaves a tamper-proof receipt. HEROS Firewall: A security guard between an AI agent and its tools that blocks dangerous calls before they run.

RouteLLM Gateway: Sends easy prompts to a cheap model and hard ones to a frontier model, so you pay less for the same answers. LLM Control Plane: One gateway in front of every LLM provider, with failover, rate limits, cost tracking and an audit log.

HEROS-RL: Trains a language model to rewrite risky database migrations into safe, zero-downtime ones. Reflexion-SWE: A coding agent that learns from its own failed attempts, without retraining the model.

GraphRAG CodeAudit: Turns a codebase into a knowledge graph so an AI can answer big-picture audit questions, not just keyword lookups. PerfBench: Hand-tuned CPU kernels for the maths inside transformer models, each checked against a reference.

Show 8 more tracked systems

HEROS Cloud: A workflow studio where the AI writes plain-English safety rules from its own failures, and a human approves them. RLVR Bias Probe: An experiment showing why a popular RL training method quietly rewards longer answers, and how a fix removes it.

EdgeInfer: Runs AI models and semantic search entirely in the browser, with no server. PeerVault: Peer-to-peer file transfer in the browser with post-quantum encryption and no server in the middle.

go-ledger: A payment service that cannot double-charge: idempotent requests and a double-entry ledger that always balances. Factory Monitor: Software that watches factory machines and cameras in real time and plugs into ERP/MES systems.

AutoPilot FDE: Finds repetitive work in a company's emails and tickets, then drafts automations for it (prototype). Mobbin UI Compiler: Generates accessible React screens from real-app design references.


04 Shipped to real users

CTrackAI: AI calorie tracker: snap a photo of your meal, scan a barcode, and track macros, fasting and water. Preeo: Cycle and health companion: log cycles and symptoms, see BBT charts, and share reports with a doctor.

CTrackAI: App Store · Google Play · case study  ·  Preeo: App Store · Google Play · case study


05 Ask my README

Click a question and GitHub opens an issue with it already filled in. Submit it, and a bot built into this repo replies in the thread, usually within a minute, using live data from my repos. I get notified too and often follow up personally.

What's your strongest project? What are you building right now?

How do you prove your code works? Are you open to relocation or remote?

How do you work with AI coding agents? What's your tech stack?

Ask anything else

💬 Be the first to ask · you can also name any repo, e.g. ask: tell me about PeerVault


06 Deep dive & credentials

Go deeper What you'll find
📚 Full engineering log Every project with its exact test counts, metrics and honest limits
🧾 Evidence matrix · Release audit What was measured, how, and what is explicitly not claimed
🧪 Case studies CTrackAI · Preeo · FieldLens
🤗 Hugging Face 11 live Spaces · 2 datasets · portfolio hub
🎓 Anthropic · Building with the Claude API verify credential
🎓 Anthropic · Claude Code in Action verify credential
🎓 Anthropic · MCP: Advanced Topics verify credential
🎓 Anthropic · Teaching AI Fluency verify credential
🎓 HackerRank · SQL (Intermediate) verify credential

This README rebuilds itself every hour from live GitHub and Hugging Face data · last change 2026-11-14 · engine: skin/build.py

Pinned Loading

  1. certiflow-ai certiflow-ai Public

    AI-powered compliance automation platform — autonomous agents verify SOC 2, ISO 27001 controls 24/7

    TypeScript 7 2

  2. ContentForge-AI ContentForge-AI Public

    Document parsing and structured content analysis pipeline.

    TypeScript

  3. cybermentor cybermentor Public

    AI-powered OWASP security learning platform — learn to hack, then fix. Built for NextDev Hackathon 2026.

    HTML

  4. fieldlens__ fieldlens__ Public

    FieldLens - AI coaching app for tradespeople built with Expo React Native

    TypeScript

  5. flux-social flux-social Public

    TypeScript

  6. vulnhunter vulnhunter Public

    AI-powered OWASP security scanner for public GitHub repos — paste a URL, get a full vulnerability report in seconds. Built for Quantum Sprint 2026.

    Python