Haladir

Haladir is a San Francisco–based applied AI company, founded in 2025 and part of Y Combinator's Winter 2026 batch, that describes itself as a decisional AI layer for logistics and, in its launch materials, as an AI product lab combining formal solvers with large language models to make AI reliable for constrained systems. The core thesis stated by the company is that LLMs are effective as formalization tools — translating an operation's rules into precise constraints and proposing tunable parameters — while deterministic solvers (SMT/SAT, MILP, OR tools, forecasting and linear-programming models) perform the actual optimization. The company applies this to logistics, supply chain, critical software, and manufacturing, and positions the output as "operational superintelligence."

The product framework, called Nomos, is described in five stages. Stage 1, data infrastructure, unifies operational data across WMS, TMS, OMS, YMS, ERP and data lakehouses, reconciling IDs, timestamps and event semantics so that SKUs, orders, shipments, lanes, pallets and shifts become first-class objects in an operational graph. Stage 2, process intelligence, uses object-centric process mining on system events to reconstruct a live digital twin of the warehouse floor, lanes and labor schedule. Stage 3 combines ML predictions (demand, dwell, ETA, no-show risk) with constraint solvers to generate decisions for vehicle routing, pick-path optimization, wave release, dock-door assignment, slotting, labor allocation and inventory positioning. Stage 4 covers implementation channels: operator review for high-stakes decisions, direct write-back into WMS/TMS/YMS/OMS for well-constrained decisions, and AI agents for multi-system coordination. Stage 5 provides monitoring and observability, including decision traceability, recommended-versus-executed comparison, capture of operator overrides, and root-cause attribution across data drift, workflow drift, forecast error and solver infeasibility. The company also markets a Nomos Sandbox demo.

Alongside the logistics application layer, Haladir builds at the model training layer, supplying solver-based reinforcement-learning environments and data pipelines to frontier AI labs. A third-party profile characterizes this work as formally verified training data with machine-checkable contracts and weakest-precondition-driven specifications, verification-aware RL environments where rewards are tied to formal specification compliance, and model validation frameworks for AI-generated code.

Founding story

The founding team states that in high school they published work cited in IEEE and Elsevier Q1 journals in operations research and machine learning, and that they left Carnegie Mellon, Princeton and the University of Virginia to build Haladir. Listed founders are Jibran Hutchins (co-founder and CEO, Carnegie Mellon), Quan Huynh (co-founder, described elsewhere as CIO), Joseph Tso (co-founder, CS at Princeton) and Preston Schmittou (founder, UVA Wise).

Business model

B2B. The company sells to enterprises running logistics operations and, separately, supplies solver-based RL environments and training data pipelines to frontier AI labs. A third-party profile classifies the customer profile as subscription SaaS plus services and consulting, with the company positioned as an AI infrastructure layer. Deployments are described as scoped per customer, with integrations, unified schema, digital twin, models, solvers and decision classes tuned to each engagement.

Not specified in the sources beyond a third-party classification of B2B with a subscription SaaS plus services and consulting customer profile; per-customer scoped implementations are described, and the company also supplies data and RL environments to AI labs.

Traction

Team size of four listed by Y Combinator (a third-party profile lists an employee range of 5), with five open roles posted in San Francisco, including a founding operations research engineer at $150K–$220K. The company states it is working today with 3PLs, distributors and frontier AI labs, and that it is engaged with one of the leading AI labs on post-training. Total disclosed funding is $4.3M. No customer names, revenue or usage metrics are disclosed in the sources.

Latest developments

March 2026: public launch post and video positioning Haladir as an AI product lab combining formal solvers with LLMs, with stated work alongside a leading AI lab on post-training. A seed round led by BoxGroup and Susa Ventures, with Sunflower Capital, Valkyrie Ventures, XPRESS Ventures and angels participating, brought total funding to $4.3M; a third-party database dates the seed round to 2026-04-29 and an earlier pre-seed to January 2026. The website currently promotes the Nomos framework and the newly introduced Nomos Sandbox, and the company describes near-term focus on improving existing 3PL ML stacks (demand forecasting, pick-path optimization, ETA prediction) before expanding to broader decision processes.