Haladir

Building Operational Superintelligence.

Winter 2026ActiveB2BOperationsReinforcement LearningLogisticsData EngineeringOperationsSan Francisco, CA, USA

Company

https://www.haladir.com

Haladir is an applied AI product lab for verifiable domains, enabling reinforcement learning & model harnesses to scale to economically-complex tasks via the formalization of the informal. We apply solver-based methods, SMT/SAT solvers, operations research, and formal verification to make RLVR and agent deployment possible in domains where correctness has never been formally defined. Just as code's internal verifiability unlocked exponential gains in AI software generation, we seek to unlock the same dynamic for larger operational use cases of AI: logistics, supply chain, ERPs and beyond.

Verdict

High Signal

Market Opportunity

Operationally complex industries — logistics, manufacturing, critical software — represent a massive enterprise TAM where current LLMs demonstrably fail on constrained decision-making. ConstraintBench itself proves the gap: frontier models score under 65% feasibility on standard OR tasks. Formal methods + RL infrastructure for post-training is a real need that foundation model companies will pay for. B2B enterprise in these verticals is a multi-billion dollar opportunity if the research can be productized.

Medium Signal

Founder Signal

All four founders left CMU, Princeton, and UVA to build — high conviction signal. Joseph Tso (Princeton, on leave) has 3 years of research at George Mason and is the most credible technical researcher. Preston Schmittou gave a lightning talk at IEEE URTC at MIT on graph attention networks and interned at NearStar Fusion. Jibran Hutchins (CEO, CMU) has a PE/finance research background — Grant Park Holdings, McLean Investment Group — not a technical founder, which is a concern for a deep-tech lab. Quan Huynh (UVA) has a near-empty LinkedIn with no verifiable credentials beyond Haladir. The YC announcement claims all founders published in IEEE and Elsevier Q1 journals in high school, but this could not be independently verified. Joshua Browder investing personally is a meaningful signal of confidence in the team.

Medium Signal

Competition

The formal constraint solver + LLM approach is meaningfully differentiated from operational AI incumbents (Palantir, o3/o4-class models, supply chain AI startups). ConstraintBench shows frontier models have a proven gap that Haladir is positioned to address. The risk is that Anthropic, OpenAI, and DeepMind are all actively researching formal methods and could absorb this direction. A partnership with a leading foundation model company would be a hedge, but that relationship is unconfirmed.

Low Signal

Product

ConstraintBench (arXiv:2602.22465, Feb 2026) is a real peer-reviewed benchmark — 200 problems across 10 OR domains, best model achieves 65% feasibility and no model exceeds 30.5% joint feasibility + optimality. Legitimate research contribution. RLFR is a self-published technical report (not peer-reviewed) on fine-tuning Qwen models with Dafny + Z3 as formal reward signal. No commercial product, no customers, no pricing, no demo. The company has 2 employees. The "working with a leading foundation model company" claim is unconfirmed and likely early-stage conversations.

OverallB Tier

Haladir has a legitimate research thesis — ConstraintBench (arXiv, Feb 2026) is a real paper proving frontier models fail badly on constrained optimization, and RLFR shows a credible technical approach to fixing it. The investors (YC, Susa, Joshua Browder personally, SV Angel) lend real credibility. The concern is execution: Jibran is a finance/PE background CEO running a deep-tech lab, Quan has no verifiable credentials, and the headline "IEEE Q1 in high school" claim could not be independently confirmed. More critically, there is no product, no customers, and only 2 employees. The gap between publishing a benchmark and selling enterprise software to logistics companies is enormous. Strong thesis and smart investors, but unproven team maturity and a very long road to revenue.