B.Tech Computer Science and Engineering at VIT Vellore (2024–2028). I like problems where you have to measure something before you can claim it: compiler optimisations, database performance, OS timers, and pulling reliable data out of messy documents.
OptGraph: Do LLVM optimisation passes interact differently depending on the program? OptGraph measures pairwise pass interactions on 129 programs (LLVM test-suite, PolyBench, cBench) and predicts them for programs it has not seen. The hypotheses were registered before the experiments and the tests are Holm-corrected. One of the four claims was not supported, and the results say so.
Python LLVM scikit-learn
FinTwin: A PostgreSQL performance bench. It runs synthetic banking workloads, records database telemetry every second, and trains models that predict latency and throughput before a workload runs. It then checks those predictions against real runs (280 runs, 383k operations).
Python PostgreSQL FastAPI React Docker
RegWatch: Turns the meeting minutes of India's drug regulator (CDSCO) into a structured dataset. LLM extraction is scored against a 200-item hand-labelled set (Cohen's κ 0.84 on the committee's decision). A fine-tuned BiomedBERT baseline predicts decisions from the application text.
Python LLMs OCR NLP
linux-hrtimer-os-project: How Linux high-resolution timers work inside the kernel, and how precisely user programs actually wake up under CPU load. Compares WSL2 with native Fedora, 6,000 measurements per environment.
C Linux kernel
Life Planner: An offline-first planner PWA with tasks, calendar, habits, a study timer and reminders. CI runs database row-level-security tests and Playwright end-to-end tests. Live demo
TypeScript React Supabase
FitCircle: A workout-streak app for a small group of friends. Two friends share a streak that only grows when both train, with dares, cover passes and push notifications. Every table has row-level security with a database test per migration. Live demo
TypeScript React Native Supabase
- AI & ML Intern, Data Science & AI Unit, Bajaj Finserv, Pune (May–Jun 2026): worked on an address intelligence pipeline for unstructured Indian addresses, covering NER-based address tagging, semantic address matching and validation. The code is confidential.
- Samsung PRISM GenAI Hackathon 2026: built the AI pipeline for the team's Smart Guided Troubleshooting Engine: query enrichment, LLM prompts and structured output schemas that turn a vague Galaxy device complaint into a step-by-step fix plan, plus the benchmark suite.
- Smart India Hackathon 2025, team lead: led the team, built the AI components (Random Forest crop recommendation and a Llama-3 chatbot for farmers) and gave the presentation.
Python, C, TypeScript, SQL · PostgreSQL, Supabase · React, Next.js, FastAPI · scikit-learn, pandas, Hugging Face Transformers · LLVM/Clang · Docker, Linux, Git