Neural logistics OS

Context

The client operated one of the densest physical networks on the planet: regional DCs, cross-docks, and last-mile nodes synchronized against a catalog that changed hourly. Planning lived in a patchwork of statistical models, tribal Excel, and vendor black boxes that disagreed with each other more often than they agreed with reality.

Leadership had already funded three “AI transformations.” Each produced a slideware win and a production miss. What they needed was not another model — it was a decision system that could absorb weather, port congestion, labor availability, promotions, and long-tail SKU cold starts without poisoning the head of the catalog.

The problem, precisely

Forecast error was not evenly distributed. The top 8% of SKUs were “good enough.” The long tail — seasonal, regional, and sparse — generated most of the dead inventory and most of the stockouts that damaged NPS. Existing hierarchical Bayes approaches collapsed under feature lag; deep nets overfit promotions; graph approaches from vendors could not meet the <90ms online join budget required for replenishment APIs.

What we built

Signal Forge designed a neural logistics OS centered on a temporal graph of products, locations, carriers, and exogenous signals. Nodes carried learned embeddings; edges carried time-aware capacity and substitution structure. Hierarchical Bayesian priors regularized the long tail so sparse SKUs borrowed strength without washing out local truth.

Architecture highlights

• Streaming feature store with online joins under 90ms p95.
• Temporal graph net + hierarchical priors for cold-start SKUs.
• Shadow mode with dual-write into existing planning tools.
• Human override console with full audit of every recommendation.
• Weekly causal evals against holdout DCs before each cutover.

The hard parts

Cold-start was the killer. Naïve transfer from popular SKUs created phantom demand in categories that looked similar in embedding space but not in buyer behavior. We introduced category-aware shrinkage and a refusal policy: when uncertainty exceeded a calibrated threshold, the system recommended safety stock bands instead of point forecasts.

The second killer was organizational. Planners had been burned. We instrumented “trust telemetry” — every override was labeled, and override rates became a first-class KPI alongside MAPE. Trust rose only when the system admitted uncertainty.

Rollout

Eleven weeks from signed charter to production for the first two regions. Shadow for three weeks, canary for two, then progressive traffic. Marketing and commercial teams were brought in early: promo calendars became first-class graph inputs so demand spikes were anticipated rather than explained after the fact.

Outcomes

Forecast error down 38% on the measured portfolio. Dead inventory reduced by an estimated $240M annualized. Stockout rate on promoted SKUs fell enough that commercial teams stopped running parallel “gut” orders. Finance adopted the OS outputs as the planning system of record within two quarters.

What we’d repeat

Ship the trust layer with the model. Treat marketing and supply as one graph. Prefer calibrated abstention over confident wrongness. And never call it done when offline metrics look good — only when regional operators stop fighting the tool.