I build AI systems that show their work.
RAG platforms · ML forecasting · full-stack products, tested and deployed

I'm a third-year B.Tech AI & Data Science student at VIPS-TC, New Delhi (GGSIPU, batch 2024–2028, CGPA 8.45).

I design, build and deploy AI applications end to end, from the database schema to the live URL. Every project below is a solo build with a live deployment, an automated test suite, and CI that runs on every push.

Experience: SDE Intern at Zomakarb AI Solutions LLP, New Delhi (Jul – Sep 2026). I built the company's internal RAG document assistant solo, end to end: document parsing, chunking and embedding, vector search with PostgreSQL and pgvector, and LLM answer generation behind a Next.js web app. I deployed it for internal use and presented it to the founders, and 3–4 team members used it for document lookups.

	SetuNER Which roads in flood-prone Assam will still be passable tomorrow? A forecasting and relief-logistics system built on live government flood reports, satellite rainfall and terrain data. Scores 110,000+ road segments, refreshed every morning by an automated pipeline Onset AUC 0.71 on a held-out monsoon season, against 0.50 for the "same as today" baseline Relief planner written as a linear program that reports which constraint is limiting: stock, trucks or reachability 319 API tests, 21 web tests, and outside-in alerting that emails me when a feed goes stale Python FastAPI PostGIS OR-Tools Next.js MapLibre Docker Live app · Code ·
	RAG Assistant Document Q&A that still works when a second team logs in. A multi-tenant platform that answers questions from your own documents and checks every citation against its source. Hybrid retrieval: pgvector similarity merged with Postgres full-text search Workspace isolation enforced by Row-Level Security on every table and the storage bucket Built-in evaluation harness that scores retrieval and answer accuracy across runs One GIN index cut keyword search from 1,716 ms to 311 ms on 20,000 chunks TypeScript Next.js Supabase PostgreSQL pgvector Gemini Voyage AI Live demo · Code ·
	Burnout Buster An early warning for counselors, before a student has to ask for help. A three-minute wellness check-in for students and a risk dashboard for the counselors who support them. Monotonic gradient-boosting models: more sleep or more support can never raise a risk score Every score is explained by the answers that pushed it up or down About 83% accuracy on 1,000 held-out students (the training data is simulated, and the app says so) 74 tests covering the model's guarantees, storage and UI flows Python scikit-learn Streamlit SQLAlchemy PostgreSQL Plotly Live demo · Code ·

I measure before I optimise. Sending scanned pages to OCR in parallel looked like the obvious speed-up. Measured on a 23-page scan, it took 252 seconds against 34 seconds one batch at a time, so the pipeline runs sequentially. Engineering decisions in RAG Assistant →

I publish the results that don't flatter me. SetuNER grades every forecast against a baseline once the outcome is known, and its live track record is on the forecast page, including the stretch where it lost. Results, stated honestly →

I write down why. SetuNER has 18 numbered decision records, including the ones that record a failure, so the reasoning behind the system outlives my memory of it. Decision records →

I make the system fail safely. Burnout Buster refuses to save a model that fails any of its 8 sanity checks. SetuNER serves the baseline whenever the model loses to it.

• Building Casewright: an AI security analyst that investigates incidents with tools, builds an evidence chain, and waits for a human before any risky action. Java + Spring Boot gateway done (Argon2 auth, rotating JWT refresh tokens, CI); the Python agent is next
• Looking for a software engineering or AI engineering internship in 2027: mohitks2609@gmail.com

_{Certifications: IBM SkillsBuild (Build an AI Agent, Craft Precise Prompts) · Anthropic Claude 101 · Forage job simulations (Deloitte Cyber Security, Tata GenAI Data Analytics)}