I'm in the final year of my B.Tech in Electrical Engineering at IIT Delhi, interested in LLMs, agentic systems and AI safety.
Currently working on temporal consistency in agentic LLMs and on multi-agent collaboration for long-horizon tasks.
An automated alpha-search pipeline for the Trexquant alpha competition: genetic-programming search over signal expressions, then verification, de-correlation and blending before submission. At the time, it got me ranked 1st globally.
A little Claude Code mod I made that shows live spend, per-model costs, cache state and context usage above the prompt (particularly useful for me, because I routed different models through my Claude Code via OpenCodex; it helped me track how quickly I was blowing through my Azure credits, and made me realise how expensive Astra on long context can be).
A custom model architecture I built to play Hangman that uses character-level CNNs. It beats BERT-based baselines by 26%, winning about 70% of games.
A practice research-paper implementation: context-aware decoding from Shi et al. (2023), implemented for GPT-2 so it sticks to the given context and hallucinates less.
Safe and data-efficient RL: I implemented PPO-Lagrangian from scratch to study reward–safety trade-offs in constrained control, and compared how offline RL methods (IQL, TD3+BC, CQL) scale with data. (Course project for AIL841: Advanced Reinforcement Learning.)