What we publish, and what we open-source.

Papers, models, and benchmarks in one index — the separate sub-pages are gone, so everything is filterable in place.

ICLR 2026 · with Google Research

TurboQuant — extreme compression for production

Two-stage quantisation reaching 3-bit zero-loss KV-cache compression, which is what makes long-context inference affordable at production scale.

paperTurboQuant — extreme compression for productionTwo-stage quantisation reaching 3-bit zero-loss KV-cache compression, which is what makes long-context inference affordable at production scale.2026
benchAgent eval harness — trajectory and outcome scoringThe suite behind the −68% RAG hallucination result: trajectory checks, a rubric-scored judge calibrated against human labels, and regression budgets wired into CI.2026
benchAgent response latency under production loadp95 under 90 ms for multi-agent coordination on live-event traffic, down from roughly 800 ms on the rules-based baseline it replaced.2025
paperPhysics-informed networks for industrial simulationConstraint-aware variational solvers for industrial flow problems, matching classical solvers at roughly 0.01% of the training data.2026
benchQuantum-inspired vs. classical training time97× reduction in training time on problems whose structure suits tensor-decomposable methods — weeks to hours, benchmarked against a tuned classical baseline.2025
benchEdge vision latency — ViT on-deviceSub-10 ms per frame for vision-transformer inference on embedded accelerators, measured at the 99th percentile under thermal load.2025