code
Software I lead or maintain, and the official implementations of my papers.
Libraries & Tools
AnyJev
Turns any LLM into a typed decision model with calibrated probabilities, training-free by default. A decision endpoint is a vLLM pooling server plus a few kilobytes of closed-form head. Cuts option-order sensitivity 0.230→0.073 and calibration error 0.240→0.095 with zero labels.
Drag or scroll sideways for the full field. Figures are the board's own, captured at v0.4.1; the live ranking moves as systems are added.
Drag or scroll sideways for the full field. Figures are the board's own, captured at v0.3.2; the live ranking moves as systems are added.
Paper Artifacts
SPECTRA
Transform coding for the LLM KV cache. A training-free, drop-in codec that rotates the cache into a coordinate system computed from its own statistics and spends the bit budget on the few channels that carry the signal.
WiSP
Routing-aware expert paging for memory-constrained Mixture-of-Experts serving. Treats low-resource MoE inference as a GPU working-set problem, paging in only the experts a workload actually reuses.