I build retrieval and LLM systems that run in production, and I like fixing what makes them slow or wrong.
Founding full-stack AI engineer at COROS AI since April 2025. I designed and built the AI platform from scratch as a set of services, and own features from system design to deployment.
- Faster retrieval: reworked semantic search and retrieval in our RAG system to get the numbers above.
- Memory that stays on topic: a semantic memory layer with topic-isolated user spaces and time-based decay, so context from one topic does not leak into another.
- Answers you can trust: automated evals with relevance scoring, and HyDE-based retrieval to cut hallucinations.
| Project | What it is | |
|---|---|---|
| 📚 | NextBook · live | Book recommendation system with a retrieval-augmented LLM pipeline. Deployed at Abul Kalam University Library, 200+ users, about 50% faculty adoption. |
| 🎯 | PinpointAI · live | Turns meeting transcripts (VTT or DOCX) into editable action items, with PDF export. Next.js, Express, OpenAI API. |
BS in Computer Science (AI and ML), NED University, expected 2027. Winner of the Data Insights Hackathon.
If you have a retrieval or LLM system that is slow, wrong, or hard to trust, I am happy to talk.