Quality-diversity evolutionary framework (MAP-Elites) for discovering diverse vulnerabilities in LLMs, presented at ICLR 2026 Workshop AIWILD. Latest: a longitudinal red-teaming study finding that safety alignment does not always improve monotonically across model generations, with evolved attack archives that transfer unevenly between them.

Designing compute abstractions for environments where power, connectivity, and time are all intermittent. Scheduling and fault-tolerance driven by orbital physics.

Building production-ready multi-agent systems for autonomous data pipeline management: an industry-agnostic framework handling data quality, orchestration, and monitoring. Current focus is the verifier. When loops run unattended over real systems, convergence is not correctness, and the check you write is what actually decides the outcome.

Reference implementations of acceleration techniques: speculative decoding, KV-cache compression, custom Triton kernels. 8.1x speedup, 88% peak bandwidth on A100. Latest: a fused MoE dispatch kernel in pure Triton, now on arXiv, reaching 89-131% of Megablocks' CUDA throughput at inference batch sizes and portable across NVIDIA A100 and AMD MI300X.

Mechanistic interpretability for production AI safety: circuit tracing for model internals and sandbagging detection, plus a new thread on what interpretability actually costs at inference time. Built a harness to test whether activation probes are cheap enough to run in production serving, against the standing assumption that they are not.