Hi, I'm Tanvir Bhathal

I'm a researcher and engineer focused on building products through applied research in language models and ML systems. I am currently pursuing my BS/MS in Computer Science and Mathematics at Stanford University, where I do research at Stanford AI Lab, co-advised by Chris Ré (Hazy Research) and Azalia Mirhoseini (Scaling Intelligence Lab). My research focuses span RL Post-Training, Scaling test-time inference, information retrieval, and building efficient and scalable ML/AI systems. Outside of research, I enjoy building agents to tackle real-world problems.
Previously, I've worked at MetaAI GenAI team on building an AI Agent Framework that handels millions of messages daily and at Brain Corp on AI solutions for their autonomous mobile robot fleet. I also worked at RapDevPro and ZibdyHealth, building iOS and Android apps for their respective products.

Things I've Built

A selection of projects spanning research, software, product development, a handful of small, older projects that were surprisingly fun.

AIProductAgents

Spielberg AI

A functional natural language video editor powered by multi-modal AI Agents. Winner of Treehacks 2025, $7k Rox AI Best AI Agent prize.

View Code

AIResearchTest-Time Scaling

Weaver: Scaling Verification for LLMs

A framework to develop weak to strong verifiers to scale LLM inference at test-time. Published at NeurIPS 2025.

View Code Paper

AIResearchAgents

Websight

A vision based web browser agent, powered by Websight-7B, a custom-trained 7B VLM.

Code Paper

AIResearchGNNDrug Discovery

GDINN: Predicting Drug Interactions using GNNs

A thorough analysis of the performance of GNNs on predicting drug-drug interactions, culminating in the development of a novel GNN for drug-drug interaction prediction.

View Paper

AIProductAgents

Marketmind

AI Agent orhcestration framework powering a real-time AI platform for market research and analysis. Won 1rst place Grand Prize at CMU's AI Valley Hackathon.

View Code

RLResearchGames

2048: Reinforcement Learning in a Delayed Reward Environment

A unified, distributional multi-step RL framework designed to directly optimize long-horizon performance on the 2048 game.

Code Paper

AIResearchRAG

ARES: An Automated Evaluation Framework for RAG Systems

A groundbreaking framework for automatically evaluating Retrieval-Augmented Generation (RAG) models.

View Code

Classic MLResearch

EmissionNet

Developing extremely successful and novel classic ML models to forecast air quality based on pollution from agricultural emissions.

Paper

AIProductAgents

Cohora: Everyone's Personal AI Secretary

An engine that enables seamless AI Agent collaboration with multi-agent, multi-modal intelligence. Real-time actions via websockets and universal tools integrations

View Code

MobileProduct

SOAR: Save On A Ride

iOS and Android app to match students together to save on ride sharing apps.

View Code

AIProductAgents

Libra AI: The Agentic Paralegal

An agent orhcestration framework, with multi-agent, multi-modal intelligence, RAG pipelines, and much more.

View Code

JavaGames

RoyalBattle

A Java GUI game showcasing OOP design through lane combat, elixir mechanics, dynamic health/bullet interactions, and AI or PvP gameplay

Code

AIProductRAG

2nd Chance

A Black Mirror inspired agent that mimics any human's texting and speech behavior, given proper texts and recordings, to help users simulate talking to them again.

View

HardwareSystemsGames

Snake on Hardware

Classic game snake displyaed on a manually constructed LED board controlled by an Arduino.

Code

Web ScrapingData Analysis

X Political Sentiment

To satiate my curiosity and address the issue of X APIs being difficult and expensive for reading tweets, I developed a web srcaping bot that determines political sentiment on the platform.

Code

BlenderGraphics3D Modeling

Blender Scene!

Leveraged computer graphics knowledge to create many 3D models in blender and placed them in a scene with proper lighting and cohesion between objects.

Code