Hemnaath Balasubramani

Smallholder farmers miss crop-insurance payouts because filing a claim usually means paperwork, an adjuster visit, and a form in a language they may not read. ClaimFarm turns a single Telegram photo into a filed claim in about a minute. A vision model reads the crop and the damage, thirty days of weather at the photo's location back it up, and the farmer gets a confirmation in their own language. I built all of it: the agent, the FastAPI backend, and a six-signal fraud check that cross-references EXIF, GPS, capture time, and near-duplicate images before anything is approved. It ran end to end for the Global AI Hackathon with Qwen Cloud, with 20 of 20 integration tests passing.

Stack Python, FastAPI, Qwen-VL-Max, embeddings, Alibaba Cloud DashVector and Function Compute.

Somerville has 2,404 combined-sewer segments and a $1.29B separation program, and the question of which pipe to dig up first is scattered across disconnected GIS layers. In one day I built BedRocked, which scores every segment from 0 to 100 for dig-readiness by fusing Cyvl's AI street scans with the city's sewer GIS. I trained a catch-basin condition classifier by distilling Claude Vision's labels into a smaller model that scored all 381 basins for cents each, then added a natural-language search over the data and an Autodesk export that renders the dig plan in 3D. Built for the Cyvl, Autodesk, NVIDIA, and City of Boston hackathon, and still online.

Stack Python, FastAPI, scikit-learn, Anthropic Claude, Autodesk Platform Services.

A co-op search means living in a spreadsheet and rewriting your resume for every posting, and most AI resume tools are happy to invent the experience the posting asked for. job.os is my fix for that. It tracks applications on a Kanban board, finds roles across 85 company job boards plus feeds that carry every employer on them, and tailors a resume against one specific job with an agent that cannot make things up. Every claim has to trace back to a verified fact, unverified numbers are stripped out of the draft before anyone reads it, and a requirement my profile cannot support becomes a question instead of a bullet. Each posting is also scored against my verified skills before I spend any time on it, and read for the things a score will not tell you: sponsorship, citizenship, clearance and export-control requirements that make a role unwinnable on a student visa. The resume itself is compiled as real Typst and LaTeX into one of six well-known templates, then held behind a quality gate: a deterministic review score, a one-page and selectable-text check, and a keyword-coverage read against the posting. Nothing downloads until it passes. Live at jobs.hemnaath.tech.

Stack FastAPI, LangGraph, Anthropic Claude, Appwrite, Postgres with pgvector, Typst and LaTeX rendering, and Sentry.

While you scroll a job board, there is no quick way to know whether a role actually matches your resume. RoleReveal is a Chrome extension that scores each posting against your resume and puts the score right on the page, working across LinkedIn, Indeed, and other sites through per-site and universal adapters. API keys sit behind a proxy with per-user rate limits, and your resume is masked before it leaves the browser. It is published and live, with a landing page at rolereveal.app.

Stack Chrome extension APIs, a Vercel proxy, and LLM scoring.