A platform that models a logistics network as a live knowledge graph and propagates shocks through it (a snowstorm at one hub, a fuel spike, a retail-sales miss) to forecast where delays will cascade, which routes are about to bottleneck, and which shipments are quietly at risk, all before it surfaces in the tracking data. Graph-context features already cut short-horizon forecast error by 7-19% on logistics, with conformalized prediction intervals for honest uncertainty. Now hardening the autonomous analyst loop that expands the graph, retrains, backtests leak-free, and writes a decision brief on its own.

Self-hosting local LLMs that pair-program and automate the day-to-day, with a small harness that stress-tests them and tracks hallucination rate per prompt template. Swapping a model becomes a measurement, not a guess.