Working paper on LLM hallucination errors using source signal, matching, reconstruction, and decoding noise.
ALGORITHMS // SYSTEMS // INFRASTRUCTURE
We design and deploy
high‑performance ML systems
for the hardest problems.
Reliable, optimized algorithms from first principles: event‑driven coordination, reactive, and conflict‑aware
. Built with precision in the open for maximum visibility.
PROJECTS
Validation lab for energy-based constraint relaxation with PSON, stability guards, and gate-benefit couplings.
Working code and reproducible experiments for PSON on sparse optical coherence and discrete phase-control tasks.
Kernel-based manifold learning with adaptive bandwidths, early stopping, and streaming embedding updates.
PyTorch monotonicity score for 1D sequences using sublinear sampling.
Small modules coordinated by a global free-energy objective, with sparse couplings and logged relaxation metrics.