research engineer · nvidia
email · github · x · google scholar
I work on post-training at NVIDIA. I'm broadly interested in what happens to models after pretraining: fine-tuning, synthetic data, agents, and the systems that tie them together.
AEVAL: From Anecdotal to Deterministic Testing for Agentic Skill Workflows
ICML 2026 · Workshop on Statistical Frameworks for Uncertainty in Agentic Systems
Skill evaluation for agents today is anecdotal: ask an agent to try the skill, watch the demo, form an impression. AEVAL replaces that with a deterministic, CI-integrated pipeline where every skill change runs against a declared evaluation contract, with the executor structurally separated from the grader so an agent cannot silently self-correct and then pass its own patched outputs. Validated on real skills in a production agentic stack across multiple agent SDKs.
NVIDIA TAO skill bank, 2026
A closed loop that trains a detector, scores it, runs gap analysis on the misses, near misses, and false alarms, then routes each gap to the data that fixes it, mined real images or synthetic defects generated onto clean parts, and retrains until the gaps run dry. Applied to object detection for automated optical inspection, where the loop matches full-data fine-tuning on a fraction of the data and keeps improving past it.