Work

In my internship with Amazon AGI, I worked on RL for a browser-use model. I led model performance on two benchmarks & worked on algorithms/performance.

As a constrained optimization problem, LLMs can use RL to invent their own compression schemes to increase its context window.

Language models, when trained on hidden-information games, naturally learn deceptive techniques to win the game by any means.

Research under Cohere Labs for a compute-efficient post training to represent different languages as modalities for multilingual language models.

Let agents actively observe the terminal and proactively fix training runs, security alerts and potential bugs.

Local tunnel MCP for text agents to interact with your OS. Poke can text others, fill out forms on your browser, play music and fix code.

Giving vision to Karpathy's nanochat for <$10 of compute, by implementing LLaVA via SIGLIP encoder injection and fine-tuning on vision Q&A.

Open-source background coding agent with 1.4k stars on GitHub. Feature-filled agent that works in a MicroVM with full codebase understanding.

Multimodal agent and long-context video understanding to help hollywood editors. Worked on the infra, video retrieval & the agent.

Tricked a Galaxy S24 to run Moondream 3B VLM locally, with quantization + local linux setup on phone. Built at TreeHacks 2025

Designed and implemented GPU-optimized voxel grids for humanoid design team in Waterloo. Co-led ML team.

A decentralized cross-device model training system with model and tensor parallelism to reduce compute needed to train large models.

Generating policy recommendations with AI agents from citizen complaints.

One of the first implementations of coding subagents to work together to solve hard, diverse coding problems.

City simulation of Los Angeles with AI Agents, simulating human behaviour and optimizing transit routing with RL.

Worked with Tempo Labs to build generative UI agents in multiple programming languages concurrently.

Self-driving car design team, WATonomous, in Waterloo. Working on a low-latency data-driven controls model for the car.

Worked with Aviato to build a recommendations system for recruiting top AI talent amongst 300k+ engineers.

Helping with the software behind Bracket Bots, a self-balancing robot for under $200 that can roam around your house.

Worked on safety simulation software for New York trains with custom network protocols.

Inpainting with diffusion language models to fill in missing text. In Progress.

256 Dimension audio embeddings for semantic audio analysis using waveforms and Fourier Transforms, trained with contrastive learning.

Long-term memory with multimodal knowledge graphs to search 7 days of video and audio within 5 seconds. Winners @ Hack the North 2023.

Built an autonomous model tank that navigates campus walking paths and delivers small parcels. Implemented path planning, simulations, object detection, PID control, and CV.

Deep learning analysis of seismic frequencies and local policy to design affordable earthquake-resistant buildings. Worked under RippleX Fellowship, RBCx.

Trained a smaller VQGAN+CLIP to generate images from text and poetry prompts. Scaled up inference to run on MacBooks efficiently.

An offline mesh network written in Swift via MultiPeer Connectivity to allow for cross-device transfer of files entirely offline, creating a chain of encrypted nodes.