Merce's existing forecast ran overnight on a warehouse-wide batch. By the time the plan reached the floor, it was already stale. The brief was simple and brutal: get a credible forecast to the routing engine inside one second of an order landing, across every site at once.
We distilled the forecast model onto a quantized runtime, then placed an inference replica in each warehouse so the decision never left the building. The orchestration layer batches low-priority queries and preempts the moment a high-priority order arrives.
They hit the latency target in week nine. We stopped arguing about the plan and started arguing about which routes to retire.
VP Operations, Merce
What broke: the eval gate failed silently for 11 hours on a Tuesday when a upstream schema drifted. We added a dead-letter audit and a hard fail-closed on eval regression, which is now standard on every B2B commission.
The constraint was the interaction budget. A spending question has to feel instant, which rules out a round-trip to a hosted model on a cold train. We distilled a 3B model to run on-device, with a cloud fallback only for the long tail of queries.
Personalization runs locally against the user's own transaction graph, which never leaves the phone. The orchestration layer adds guardrails for regulated advice and refuses to draft anything it cannot source.
It answers in the time it takes the keyboard to close. Our users stopped noticing the AI, which is the only review we wanted.
Head of Product, Orbital
What broke: the on-device model misread a recurring transfer as fraud for a subset of locales. The fix was a locale-aware eval set in the training mix, and we now ship per-region eval gates on every consumer build.
Halden's catalog had outgrown its search vendor, its ranking was a static rulebook, and checkout was a separate contract entirely. The brief was to collapse all three into a single engine that could reason over inventory, intent, and margin in real time.
We rebuilt the data plane first, then trained a ranking model on eighteen months of real funnel data, then added an agentic checkout that can hold a cart together across devices and payment failures. One stack, one source of truth.
Conversion is up a third and we cut two vendor contracts. The engine pays for itself every quarter now.
Director of Digital, Halden
What broke: the ranking model over-optimized for margin and starved the long-tail catalog of impressions. We added an exposure-balance term to the loss function and a daily diversity report, both now standard on commerce work.
Your system,
next on the bench.
Send us the problem. Two business days later you get a scope, a model shortlist, and a deployment plan. No decks, no theatre.