Projects

Built an agentic inference pipeline solving olympiad-level mathematics problems using a locally-served 120B-parameter language model with tool-augmented reasoning, as part of the Kaggle AI Mathematical Olympiad Progress Prize 3 (AIMO-3).

• ·Designed the end-to-end architecture: vLLM model serving with mxfp4 quantization and fp8 KV caching on a single H100 GPU, paired with a persistent Python sandbox enabling symbolic computation and numerical verification.
• ·Implemented a 5-component weighted-entropy aggregation system scoring parallel solution attempts by token-level confidence signals, outperforming naive majority voting on stochastic reasoning tasks.
• ·Placed in the top 4% of global competition participants.

Developed and optimized a reinforcement learning agent for strategic decision-making in a multi-agent Pokémon battle environment at the NeurIPS 2025 competition.

• ·Implemented policy optimization, reward shaping, and environment-specific state encoding to improve sample efficiency and generalization across matchups.
• ·Minimized inference latency while preserving strategic depth — balancing exploration and exploitation under strict runtime constraints.
• ·Achieved Top 10 placement globally through iterative experimentation, hyperparameter tuning, and systematic failure-case analysis.

Built an ML model predicting soil condition changes using beetle population data as ecological indicators of climate shifts, in collaboration with the NSF HDR National ML Challenge.

• ·Led the modeling and feature engineering pipeline, integrating ecological datasets with supervised learning to capture nonlinear species-environment relationships.
• ·Focused on cross-validation robustness, interpretability, and generalization across geographic regions.
• ·Won 1st place at UCSD-level competition (~200 participants) and placed 6th nationally.

Multi-agent system for complex healthcare question answering, routing questions across specialized agents (Pharmacology, Diagnostics, Treatment Planning) with domain-specific GPT-4o prompting.

• ·Achieved 84.5% validation accuracy through agent orchestration, structured outputs, and error-driven refinement.
• ·Implemented dynamic routing, domain-specific prompts, and cost-efficient model selection (GPT-4o / GPT-4o-mini).
• ·Placed Top 25% of participants globally in the Harvard Medical School-hosted competition.

Built an identity-aware AI agent that connects to Gmail and Google Calendar, learns the users communication style from sent emails, plans tasks, and drafts replies in the users voice — with configurable approval and auto-send rules governing every action.

• ·Designed the full agent architecture: delegated OAuth access to Gmail and Calendar, a style-profiling layer that extracts writing patterns from sent mail to ground reply generation, and a planner that maps inbox and calendar context into prioritized task sequences.
• ·Built an approval rules engine defining when the agent can auto-send vs. hold for review, with domain-level and length-based controls — ensuring autonomous action stays within explicit user-defined boundaries.
• ·Implemented an activity timeline for full traceability of every agent decision: planning, drafting, rule evaluation, and action outcome.

Quiz-style web app where users write a tweet and guess which celebrity they sound like — using NLP classification with interpretable predictions.

• ·TF-IDF vectorization + Logistic Regression to classify tweets across multiple celebrities.
• ·X (Twitter)-style interface reveals identity only after prediction, creating a game loop.
• ·Explains predictions with keyword contribution breakdowns for transparency.