Research
The lab works 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: confident agents can still disagree, and an orchestrator fusing their evidence can be quietly wrong. We estimate system-level uncertainty across heterogeneous agents, detect errors and conflict, and decide when to abstain or escalate, so agentic perception is trustworthy enough to field.
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 shifts with model scale, family, and architecture. We treat 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 theme across the lab: don't trust the headline number. We build 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 deploy in medicine and defense.
An orchestrator pools the semantic opinions of several heterogeneous vision-language agents to estimate system-level uncertainty, separating each agent's own doubt from disagreement between agents, for error detection, conflict diagnosis, and abstention.
A multi-agent system that decomposes complex spatio-temporal video questions into tool-augmented, multi-step reasoning (active perception, tracking, temporal retrieval, hierarchical memory, and symbolic verification) under edge 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 and family.
A pre-registered result showing the KL distillation term acts as a stability anchor rather than a knowledge-transfer channel. A self-anchor control transferring zero information reproduces essentially all of the benefit.
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 .