Research
My work runs along three connected lines: perception you can trust, models small enough to deploy, and the measurement discipline that keeps both honest. A recurring principle ties them together: don't trust the headline number.
Multi-agent systems built on vision-language models fail in ways single models don't: individually confident agents can still disagree, and an orchestrator that fuses their evidence can be quietly wrong. I work on estimating system-level uncertainty across heterogeneous agents, detecting errors and conflict, and knowing when to abstain or escalate, so agentic perception is trustworthy enough to deploy.
Getting a capable model onto constrained hardware is a routing problem: when to distill, when to prune-and-recover, when to quantize, and how that answer shifts with model scale, family, and architecture. My group studies compression as a predictive measurement science, down to running distilled, quantized models on GPU-less edge CPUs with a measured size, speed, and accuracy envelope.
A recurring theme across my work: don't trust the headline number. I develop uncertainty quantification, calibration, and test-and-evaluation methods, including statistically valid accept and stop rules for autonomous experimentation, controls that separate genuine gains from evaluation-format artifacts, and assurance methods that make AI safe to field. This is the foundation my agentic and edge work is built on.
An orchestrator dispatches a visual query to several heterogeneous vision-language agents, then pools their semantic opinions to estimate system-level uncertainty, separating each agent's own doubt from disagreement between agents, so it can detect errors, flag conflict, and abstain when the collective isn't sure.
A multi-agent system that decomposes complex spatio-temporal video questions into tool-augmented, multi-step reasoning (active perception, detection and tracking, temporal retrieval, hierarchical memory, and symbolic verification) under edge latency and compute constraints.
Turns compression-route selection into a predictive science: when elastic-width search beats structured pruning-with-recovery beats post-training quantization, and how the optimal route shifts with model scale, family, and architecture.
A pre-registered result showing the KL distillation term acts as a stability anchor rather than a knowledge-transfer channel. A self-anchor control that transfers zero information reproduces essentially all of the benefit.
Accept and stop rules for agent-driven experimentation that keep AI 'scientists' from chasing noise, showing a previously reported large effect was a small-sample artifact.
Uncertainty-aware models for detecting rare leukemia cells (measurable residual disease) in flow cytometry, with an agreement-and-triage rule that auto-decides easy cases and routes disagreements to a pathologist.
Several results are under peer review or subject to public-release review; they're described here at the capability level. For the complete, dated publication record, see Google Scholar .