Benchmark on browser-use agents interacting sites with dynamic, stateful, and temporal behaviors driven by asynchronous requests.
Research & Projects
research
Agentic AI orchestration system with a staged reasoning chain to produce structured reports (SoTA on PubMedQA, accepted to RSNA).
built
Remote GUI automation layer for computer-use agents; mirror a local desktop's apps and files, then run tasks concurrently, with no desktop takeover.
Bitcamp track winner. Multi-agent orchestration pipeline that turns documents into narrated 3Blue1Brown-style videos with Manim and ElevenLabs.
Former CTO, built Full-stack college application platform with >5k users and >100k in sales.
Reinforcement learning environment and agent for the card game Scoundrel, trained with MaskablePPO and potential-based reward shaping.
Reimplmentation of Pangram's technical report; trained AI-generated text detector using iterative hard negative mining. Outperforms GPTZero on checkforai benchmark.
Digit classifier with spring-coupled gyroscope network, made trainable via backprop through a differentiable ODE solver. Inspired by unconv.ai blog.
open source
Qwen3-4B finetune on synthetic chest CT radiology reports with QLoRA SFT and GRPO, using a local judge model for reward scoring.
Contributed Gemini provider support and deterministic verifiers to Harbor's unified computer-use agent (upstreamed from an internal repo).