Experience & Education

I own the AWS AI/ML skill in AWS’s open-source agent toolkit. I built its evaluation infrastructure on promptfoo, an end-to-end simulator with Strands Agents, and the evaluation criteria used across the skill catalogue.

Built multistep KYC benchmarks with evidence-backed follow-up questions, and persistent onboarding memory with structured state, so long-horizon decisions stay auditable and can resume from incomplete evidence.

Reward-hacking detection, evaluation of auto-research agents, and the Agent-Native Research Artifact repository.

Terminal-Bench-Science and Harbor contributions; atomic-scale TEM analysis: segmentation, denoising, tracking and TensorRT inference over 10K+ frames.

Automated SageMaker Studio UI canary and smoke testing with an AI testing framework on Bedrock, and deployed a customized browser agent for UI automation on a CloudFront–ALB–EKS stack.

Built sliding-window U²-Net inference that keeps ultrasound segmentation stable across frames, and a PyQt multi-anatomy annotation framework used to produce and correct the training data.

Advised by Prof. Hao Chen. Dissertation: RepoGenReflex.

ContractBench, AgentECC, ForkLeft, FuzzAug.

Attention-head Shapley attribution for BERT via neuron-Shapley, applied to intent recognition.