The San Francisco startup pairs old-school optimization solvers with large language models so software can reason through hard constraints - not just sound confident. The pitch to logistics operators: keep your systems, add a brain on top.
Ask a modern AI model to write a wedding toast and it will hand you something charming in three seconds. Ask the same model to schedule four hundred trucks against a hard delivery window, a labor cap, and a dock that can only fit twelve trailers at once, and it will still answer in three seconds. The difference is that the second answer is often wrong, delivered with exactly the same confidence as the first. Haladir, a San Francisco company in Y Combinator's Winter 2026 batch, was built around that specific failure.
The founders describe their work in a phrase that sounds abstract until you watch a distribution center try to reroute ten thousand orders during a storm: operational superintelligence. Stripped of the buzzword, the idea is narrow and practical. Most of the decisions that move the physical world - which pallet to pick first, which carrier gets the load, which worker covers which shift - are constrained optimization problems. There is a right answer, or at least a measurably better one, and getting it wrong costs money. Language models are fluent at these problems and unreliable at them. Haladir's bet is that fluency was never the point.
The phrase Haladir keeps returning to is that leading models are “confidently incorrect.” When a system is asked to optimize across millions of interdependent variables under hard constraints, a pure language model tends to produce something that reads well and does not hold up. It understands the words in the prompt - the semantic context - without understanding the constraints underneath them. A truck cannot be in two cities. A shift cannot run negative hours. A warehouse aisle has a width.
In logistics, a wrong answer that sounds right is worse than no answer at all, because someone trusts it. That is the gap Haladir is trying to close, and it is why the company talks less about raw capability and more about verifiability - decisions an operator can actually check before acting on them.
Haladir's technical move is not to make the language model smarter. It is to hand the hard part to a different tool. The company combines large language models with formal solvers - the operations-research machinery that has quietly optimized airline schedules and factory floors for decades. The solver handles the constraints. The model handles the reasoning, the interpretation, and the messy human context around the decision. Together they turn an unreliable output into what the company calls a verifiable, optimal decision.
This split matters because it explains why Haladir exists now rather than ten years ago. Solvers have long been powerful and painful to deploy - rigid, expensive to configure, hard to connect to live operations. Language models made them approachable. The founders were circling this exact intersection long before it was fashionable; while still in high school they published operations-research and machine-learning work cited in IEEE and Elsevier journals. The timing, not the interest, is what changed.
The product framework carries a fitting name: Nomos, the Greek word for law or order. It is built to sit on top of a company's existing stack rather than rip it out - a deliberate go-to-market choice that lowers the barrier to a first deployment. It runs in five stages, from raw data to a decision someone can trust.
The concrete decisions Nomos targets are the unglamorous ones that run a warehouse and a fleet: vehicle routing and pick-path optimization, wave release and dock-door assignment, labor allocation and inventory positioning, and coordination across multiple carriers at once. None of these make headlines. All of them break when they go wrong.
There is a second business hiding inside the first. Alongside the logistics application, Haladir builds solver-based reinforcement-learning environments and data pipelines for frontier AI labs. In plain terms, they help train models to satisfy constraints during post-training, then turn around and use that same capability to serve operators. The company has said it is working with one leading AI lab to integrate its technology into that lab's post-training process.
The two layers feed each other. Better constraint-trained models make Haladir's decision agents stronger; real operational deployments generate the environments and data the training side needs. It is a tidy loop, and it is unusual for a company this young to be standing on both sides of it.
Where Haladir fits is easiest to see by what sits on either side of it. On one side are the classic optimization vendors - the solvers and operations-research toolkits that are precise but hard to deploy and blind to context. On the other are the LLM and agent startups that are easy to deploy, great with context, and unreliable on hard constraints. Haladir is aiming for the seam between them: solver-grade rigor with model-grade flexibility.
Illustrative positioning based on the company's stated approach.
The customers Haladir names are the ones for whom these decisions are the whole business: third-party and contract logistics providers, brand-owned and independent distributors, and the shippers and cargo owners who live and die by on-time delivery. For them, a few percentage points of routing efficiency is not a demo - it is margin.
Haladir launched publicly through Y Combinator on March 15, 2026, and its $4.3M seed round followed with BoxGroup and Susa Ventures leading. Sunflower Capital, Valkyrie Ventures, XPRESS Ventures, Y Combinator and SV Angel joined, along with angels including Joshua Browder, whom the founders called their first believer. One investor reading the round noted that the BoxGroup-and-Susa combination reads like a bet on product velocity as much as on the vision - the sense that the hard architectural questions about what “operations” means at scale are arriving fast, and the team is expected to answer them quickly.
The company was founded in 2025 by Jibran Hutchins, who serves as CEO, along with Quan Huynh, Preston Schmittou and Joseph Tso. They came out of Carnegie Mellon, Princeton and the University of Virginia, with backgrounds concentrated in operations research and machine learning. The team is small - around five people - and research-heavy, which fits a company whose product is essentially a claim about how decisions should be made.
What makes the founding story worth noting is not the pedigree but the persistence. The same people who published constraint-optimization research as teenagers are now trying to make that math deployable in the one place it has always been hardest to trust: live, high-stakes operations where being wrong is expensive and being confidently wrong is worse.
Whether Haladir becomes the standard decisional layer for logistics or one interesting attempt at a genuinely hard problem is still open. The company is months old. But the framing is clear-eyed in a market full of AI that impresses in a demo and disappoints in production. Haladir is not promising a smarter chatbot. It is promising a decision you can check - and in a warehouse at 2 a.m. during a storm, that is the only kind worth having.