Rigorous engineering for the AI frontier.

Where the math has to be right and the code has to ship: AI-safety research, accelerator and hardware tooling, and the research software serious labs depend on.

Selected work.

Client systems, governance positions, and the ecosystem infrastructure we maintain.

AI Safety

ARIA

Multi-year UK government Safeguarded AI work, from the formal methods themselves to the software that makes them usable.

AI Safety

MIRI

Haskell infrastructure and upstream compiler work for foundational alignment research.

AI Hardware

Groq

Software engineering, infrastructure, and tooling for AI accelerator teams.

Research

MIT CBMM and Kanwisher Lab

Research software and human-in-the-loop behavioral data collection.

Hardware

Myrtle.ai

DDR4 memory-controller IP in Haskell for FPGA-based neural-network deployment.

AI Safety

Formal-methods software

Research tooling that makes frontier safety methods executable, inspectable, and usable by technical teams.

Why us

Obsidian works fluently across strong type theory, category theory, formal methods, and production software delivery. That combination is why serious AI organizations bring us in not just to implement the research, but to help produce it.

Across the AI stack

Our work spans AI safety, Haskell infrastructure for alignment research, AI accelerator tooling, research software at MIT, FPGA-adjacent Haskell systems, and Nix infrastructure for code execution.

AI where correctness isn't optional.

Discuss AI safety, research infrastructure, or high-assurance AI tooling.