I build with LLMs, knowledge graphs, and a worrying amount of caffeine. Just wrapped an SDE internship at 123 of AI — cohort learning, payments and a recommendation engine on Azure. Before that: real-time pipelines for a defense-intelligence project, and a first-authored paper on ontology-guided knowledge graph extraction.
I'm a final-year AI & ML student at BIT Mesra. Most recently an SDE intern at 123 of AI, shipping a cohort-learning platform, a Razorpay monetization stack and an embeddings-based recommendation engine on Azure. Before that, Kafka pipelines and Neo4j graphs for a government intelligence project at Konect U.
By night, I either ship side projects, grind LeetCode, or argue with myself about which font to use on my portfolio. (See: above.)
Things I think about a lot: how to make LLMs actually useful, how knowledge graphs can replace half the SQL we write, and what a truly cognitive AI coding agent looks like — one with persistent memory and a critic loop, not a stateless chatbot. (I built one. It's below.)
I like clean APIs, dense code, and a well-placed em dash.
Ontology-guided two-pass extraction with a locally hosted Qwen3.5-9B over a Kafka document stream. Live ontology-slice retrieval cut catalog overhead by ~94%, and a six-algorithm deduplication + embedding-resolution pipeline raised search recall from 70% to 95% — with zero false merges.
feat(api):… line. Filters binary diffs to keep tokens lean.B.Tech · Artificial Intelligence & Machine Learning
CBSE · Class XII · Jharkhand
CBSE · Class X · Jharkhand
Across LeetCode, Codeforces, GeeksforGeeks and CodeStudio. Yes, I have a spreadsheet. No, I won't show it to you.
Secured a top-5 finish at the IEEE Capture-The-Flag — solving real-world security challenges in a team of four, mostly on coffee and dread.