Agastya Seth

An orchestrator that dispatches task agents across 20 to 30 hour EDA runs, and an agent chat engine used by about 800 engineers a month.

A conditional VAE that rebuilds motion-corrupted PPG before CNN and Transformer models predict blood glucose. It cut MAE from 29.6 to 25.0 mg/dL and recovered about 2,000 unusable records.

Led bringing ReNO's reward-based noise optimization into Pyramid-Flow. It sharpens single-step image models but blurs hierarchical video ones, and the write-up explains why.

Multi-agent LLMs that generate and detect synthetic hardware vulnerabilities, and a focus session on agentic test generation.

A systematic look at how LLM safety defenses trade off real protection against over-refusing harmless requests.

A PPO agent that transforms ECG data to resist re-identification linkage attacks while keeping it accurate enough for classification.