AI software engineer (intern) at Humantic AI, building production LLM features in Python for B2B sales teams. Final-year B.E. CSE student, graduating April 2027.
- LLM systems in production. Retrieval-augmented generation over SEC filings and trusted news, a central LLM router behind 72 call sites, and model choices made by evaluation: five frontier models compared on 541 real briefs for cost, quality and hallucinations.
- Reliable backends. Led Signals V2 (AI-curated account alerts for about 256 sales users) to production with idempotent retries, deduplication and source-quality filters, plus an HMAC-SHA256 signed webhook dispatcher customers integrate with.
- Optimisation over guesswork. My final-year project, MindMesh AI, forms student teams with a CP-SAT solver and uses an LLM only to read the teacher's rules and explain each team. Zero constraint violations across 60 solver runs; not one of 23 direct-LLM runs produced a valid partition.
- MindMesh AI: CP-SAT team formation from personality and skill data; the LLM only reads the rules and explains the teams.
- llm-judge: compare LLMs on your own tasks with pairwise LLM-as-judge (order-swapped to catch position bias), hallucination checks against a source, latency and cost.
- hooksign: signed webhooks for Python with replay protection, zero-downtime secret rotation, safe retries and receiver-side dedupe.
Python, FastAPI, TypeScript, React, Node.js, C# / ASP.NET Core, MongoDB, MySQL, AWS (S3, SQS), Docker, OpenAI, Claude, Gemini, Groq, MCP, OR-Tools
- President, Blockchain Club
- 1st place, National Science Day (AI cell phone detection with YOLOv3)
- Oracle AI Vector Search Certified Professional