$4.3M · AI + Logistics

What they're building

Haladir is doing two things simultaneously, and understanding both is what makes the company unusual. The primary product is an operational AI platform for 3PLs and distributors. Three layers: Substrate ingests and normalizes data across every system a modern warehouse runs on (WMS, TMS, YMS, OMS, LMS, IMS, ERP, WOS, EDI, WCS/WES) and exposes it as a single queryable operational graph, with SKUs, orders, shipments, dock doors, and labour as first-class objects. Operator deploys event-driven AI agents with actual execution authority over fulfillment waves, pick-and-pack direction, dock assignment, and exception handling, bounded by guardrails operators define. Engine runs solver-grade operations research in production: vehicle routing (VRP, CVRP, VRPTW), multi-echelon inventory optimization, mixed-integer programming, demand forecasting, ETA prediction, pick-path optimization, and shift scheduling. The secondary product is the more intellectually interesting signal. Frontier AI labs can license Haladir's substrate and engine as RL training environments, post-training corpora, and evaluation harnesses for models that need to reason about the physical economy. Haladir calls this RLFR: Reinforcement Learning from Formally-Defined domains. The thesis is that just as code's internal verifiability unlocked exponential gains in AI software generation (because correctness could be formally checked), the same dynamic can be unlocked for logistics and physical operations once those domains are formally specified. Haladir is doing the formalization work that makes logistics a verifiable domain. Once it is verifiable, RL can scale inside it. The lab sells operational intelligence to 3PLs and sells the resulting RL environment to labs like the one David Silver just raised $1.1B to fund. Both customers are real, both are paying, and they compound each other.

Why this matters

Global logistics moves roughly $10 trillion in goods annually and runs primarily on human judgment, fragmented software systems, and reactive exception-handling. The average 3PL operates 8+ disconnected systems that do not talk to each other in real time. When a shipment is delayed, a dock is blocked, or an inventory discrepancy surfaces, a coordinator is paged. That coordinator looks at multiple screens, makes a judgment call, and executes it manually. At scale, across hundreds of locations and thousands of daily exceptions, this is where margin bleeds out. AI has been able to analyze logistics data for years. What it has not been able to do is act inside live operations with the formal guarantees that operations teams require. Haladir's solver-grade optimization layer provides those guarantees. The combination of unified data, autonomous execution, and formal optimization is not a feature set that existing WMS or TMS vendors offer. It is a new layer. The AI training angle is equally well-timed. Reinforcement learning from human feedback (RLHF) has been the dominant alignment technique for LLMs. The next frontier is RL that generalizes beyond conversational domains into physical, economically complex environments. Code generation worked because code is verifiable. Logistics, with formally specifiable constraints (vehicle capacity, time windows, inventory bounds, labor rules), is the next verifiable domain. Haladir's platform creates the environment in which that generalization becomes possible.