Selected Work
A selection of things I've built, spanning AI agents, vision-language reasoning, medical imaging, and applied ML.
A security crew that catches a compromised in-policy agent inside an invoice-approval workflow. A deterministic five-rule gate re-derives the legitimate payee from company records and blocks payment redirects and amount tampering, with HMAC-signed authorization and tamper-evident audit logs. Validated against AgentDojo and InjecAgent attacks.
A local companion for Claude Code: watch the agent work live (phase, streaming answers, and the tool feed) instead of doom-scrolling during the wait, and X-ray where your tokens actually go.
Zero-shot vision-language reasoning via a Dual-layer Scene Graph Chain-of-Thoughts. Structures visual scenes into layered scene graphs to guide step-by-step reasoning in VLMs, accepted to the AAAI 2026 Student Abstract Track.
Temporal Weakly-Supervised Triplet recognition in surgical videos, identifying <instrument, verb, target> triplets from weak labels using temporal cues. Accepted to the AAAI 2026 Student Abstract Track.
A persistent memory layer for LLMs that stores, retrieves, and de-duplicates facts across sessions so agents can remember context over time. An exploration of the open problems in agent memory: conflict resolution, forgetting, and temporal reasoning.