I build AI stuff and anything else I find interesting, RAG pipelines, multi agent systems, Full Stack apps, and generally anything that looks shiny at 2AM in the morning :)

Final year Computer Engineering at Thapar, with an IIT Madras Diploma in Programming under my belt, because apparently one courseload wasn't enough

Right now I split my time between two internships: building production AI systems at Oddmind Innovations, and working on accent invariant speech research with Samsung PRISM

I like building things, breaking them, and occasionally wondering why I decided to build them in the first place

Almost forgot, Welcome to the site! Leave a note if you like it :D

aahilminookhan@gmail.com

The whole thing, as a page you can scroll.

Open the site → Download PDF ↓

Server-rendered, and works with JavaScript disabled.

One message away.

aahilminookhan@gmail.com

2,958 contributions in the last year.

A snapshot, short and sweet.

Last updated August 2026

Borrowed from Derek Sivers, who has been telling people to keep one of these for years.

It behaves like a desk. Nothing here is a page.

Your first time here. This line is counted in your browser, not on the server.

Leave something behind.

Four small things to do instead of reading a résumé.

2026 · System · NextJS, FastAPI, LangGraph, pgvector

A compound risk engine that catches the hazard no single sensor is looking for.

1 / 3

• Fuses gas readings, work permits, isolation state and worker location into hazard pathways — the danger that only exists when four boring facts line up.
• Cut the false negative rate from 44.5% to 0% across 593 statutorily labelled cases, and alarms 28 minutes before a conventional single sensor threshold fires.
• LangGraph multi agent assessment pipeline with conditional agent fan-out and hybrid pgvector/SQL retrieval over a clause level regulatory corpus.
• The risk verdict stays fully deterministic, and every generated citation is validated against the evidence it was retrieved from — no confidently invented regulations.
• Durable Postgres job queue plus a non-blocking WebSocket layer with per client backpressure, so a slow browser never stalls the pipeline.
• Hash chained, tamper evident audit trail: edit or reorder a recorded safety decision and the verification endpoint names the exact sequence where it happened.

2025 · System · TypeScript, Python, AWS

Persistent memory for coding agents, so they stop re-learning your codebase every session.

1 / 2

• Gives coding agents durable project context via the Model Context Protocol.
• A VS Code extension and MCP server for context logging, project search, recent changes and branch-aware RAG retrieval.
• Event-driven async AWS pipeline that reads git diffs with Bedrock models and pulls out the decisions, risks and tasks hiding in them.
• Branch-aware RAG with DynamoDB caching: 1.26s median push-to-search latency, 1.60s p95.

2025 · Product · Local-first

A Chrome extension that turns your browsing into a private, on-device knowledge base.

1 / 3

• Local-first: nothing ever leaves the machine. The semantic embeddings run on-device.
• Multi-stage retrieval pipeline for private, low-latency search.
• Knowledge graph linking pages by semantic, temporal and contextual relationships.
• Sessionization with context-boundary detection, so your tax admin and your holiday research never end up in the same context.

2025 · Research · NL → SQL

Grounded insights from natural-language analytics — ask in English, get a real query.

• Multi-model NL→SQL pipeline for conversational data analytics with structured outputs.
• SSE streaming with fast-path routing (snapshot + cache restore), which buys both latency and transparency.
• Context-aware multi-turn handling using structured conversation state, so "and for last quarter?" still knows what you meant.
• Schema-correction retries and resilient API/SSE error handling keep the workflow reliable end to end.

Samsung PRISM · Remote · Jul 2026 – present

Teaching multilingual SpeechLLMs to cope with accents they were never really trained on.

• Work-let: accent-invariant representation learning for SpeechLLMs, mentored by Dr. Spandan Dey and Hirak Mondal.
• Reproducing and benchmarking baseline multilingual pipelines: SALM, Qwen-Audio, SALMONN, then running accent failure analysis to find which accents actually break, and why.
• Building parameter-efficient adaptation using LoRA adapters, accent-aware front-end encoders and MoE dynamic adapter routing to improve ASR/AST on accented and low-resource speech, without retraining the whole model.
• Closure deliverable is a peer-reviewed publication, so the bar here is a paper rather than a demo.

Oddmind Innovations · Bangalore, remote · Apr 2026 – present

Building an AI interview platform end to end — frontend, backend and the realtime layer.

1 / 2

• Shipping across the frontend, backend and LiveKit service: custom interviews, realtime sessions and recruiter workflows.
• Built audio interview evaluation pipelines on AWS S3 and the Gemini Files API, storing structured scores, recording metadata and recruiter-facing report data.
• Added versioned prompt templates with centralised resolution logic, so changing a prompt stops being a find-and-replace across the codebase.
• Recruiter pipeline tooling with server-side filtering and candidate workflow management.

Datesheet Generator · Thapar University · Mar – Jul 2025

Re-architected an exam scheduling system that couldn't survive a page refresh.

• Moved from session-based state to a persistent MySQL-backed workflow — the core fix.
• Secure authentication, multi-user support and savepoints, so partial work resumes instead of starting over.
• Improved the generation workflow, cutting scheduling errors by 40%.
• Integrated with Oracle PeopleSoft to automate structured data ingestion and export between the legacy system and the new one.

Thapar EduTube · Patiala · Oct 2024 – Jan 2025

Led 3 engineers building a lecture archival platform for 15,000+ students.

1 / 3

• Backend APIs and dashboards for lecture uploads, metadata management, access control and search.
• Built a CLI sync tool for local lecture management with automated YouTube uploads and diff-based updates, which took a large bite out of upload times.
• Evaluated Elasticsearch against PostgreSQL full-text search and migrated the search infrastructure — less operational complexity, same performance.
• Used k6 load testing and deployment tuning to hold up under concurrent traffic on college-hosted Linux servers.