AI and data engineer in Bengaluru. I build retrieval and analytics systems that cite their sources and refuse when they can't, and the products around them. Resume Roaster and Truly are live.
Reciprocal rank fusion, k = 60score = Σ 1/(60 + rank)Doc C lands 4th, outside the top 3.
A resume scanner I built and deployed. The code implements spatial PDF parsing, a bullet rewriter that avoids repeating a verb, and ATS-safe export. Next.js, Supabase and the OpenAI API.
Open the live siteRead the case study
Screenshot shows the product's own sample resume, not a real user's.

On four unanswerable questions it hallucinated 0 times; the plain hybrid baseline did 4 times. On 24 answerable ones it scored 45.8% where the baseline scored 60.4%, refused wrongly 12.5% of them, and took 17.7 s a question against 1.86 s for a single pass.Agentic RAG Knowledge Assistant, open source, Python. Held-out results from the 32-question test set of a 40-question dataset (8 development, 32 held-out), frozen before later auditor fixes. The sample is small: 24 answerable, four unanswerable and four false-premise questions.

A social platform built around structured 60-minute video circles of eight people and one host. Seat capacity is checked atomically inside a database transaction. Next.js, Drizzle and PostgreSQL.