Building systems
that think and act
on their own.
Final-year Computer Science student at Theem College of Engineering, Mumbai. Every project I have shipped is an autonomous agent — a drone swarm that defends airspace, a code reviewer that watches every commit, a data analyst that refuses to invent numbers, a RAG pipeline that second-guesses itself. The pattern is not accidental.
Autonomous drone defense system built for Bharat Electronics Limited. A decentralised Hunter-Killer algorithm lets drone swarms detect, prioritise, and neutralise adversarial threats without any central controller — validated across 100+ combat scenarios.
AI code guardian that hooks into GitHub via webhooks. Every push triggers Llama 3 70B analysis — bugs are flagged, tests suggested, docs drafted and posted as commit comments automatically.
The LLM is not the analytics engine. A planner emits a schema-validated JSON plan; a deterministic pandas/scipy/DuckDB engine runs 15 fixed operations, and every headline number is independently recomputed through a second arithmetic path before it is shown. No model-written code, no invented numbers — 20/20 golden-question cases pass, and 250K rows profile in 2.0s.
Corrective RAG that doesn't trust its own retriever. Dense FAISS and sparse BM25 are fused with RRF, then an LLM judge scores every chunk before it reaches the generator — strong matches go straight to generation, weak or mixed ones trigger a multi-turn agentic web search. A sentence-level refiner filters every path, and the system says "I don't know" rather than hallucinate when context comes up empty.
Fine-tuned Microsoft Phi-3-mini (3.8B) using QLoRA to adapt response style by user persona — Student versus Expert. 67% adaptation rate, 52% factual accuracy. Entire pipeline runs on a free T4 GPU at zero cost via 4-bit quantisation.
I am a final-year Computer Science student at Theem College of Engineering, graduating in May 2026. I have spent the last two years obsessing over one question: how do you make software that can reason, plan, and act on its own?
Every project I have built is an answer to that question from a different angle — defence systems, developer tooling, data analysis, retrieval-augmented generation, and language model alignment. The common thread is agency: systems that do not wait to be told what to do next.
I work best where the problems are hard and the feedback loop is short.