Open the experiments.
Show the work.

Our R&D floor explores how AI-built systems can become more explicit, inspectable and useful. Some of the work is public and open source. Some is still taking shape.

Visit the R&D floor

Turn architectural intent
into a check.

Blueprint Conformance Engine turns architectural rules you have written down into checks that run against a codebase in CI. Set up as a CI gate, it fails the check when it detects a breach of a supported declared boundary, so the rule doesn’t depend on a reviewer remembering it.

The project is public under the Apache-2.0 licence, with tagged releases, documentation and the clean and violating fixtures that show how it behaves.

Blueprint Conformance Engine

Public · open source · Apache-2.0

Input
A declared architectural rule
Check
Code against the rule, in CI
Evidence
Passing and violating fixtures
Release
v0.5.0, September 2026
An inspectable mechanism, not a productivity claim.

From research
into the products.

The same idea runs through the products: write the rule down, then let code enforce it.

Rules as code

Odin’s engineering rules compile into action guards that check an agent’s actions against registered patterns and deny prohibited ones before they run. They are pinned by hash and generated for both Claude Code and Codex. Harness packages them for your repositories.

Harness

Answers that can refuse

Odin Next builds answers from what it retrieves. If nothing relevant is found, it says so, and it refuses an answer that cites none of the retrieved sources.

Odin Next

Questions
we’re working on.

These are research directions you can explore with us. They are not self-serve product releases.

Factory Intelligence

How do quality, cost and improvement relate in AI-assisted engineering, and can a high-quality piece of agent work become organizational evidence?

Research direction

Odin Gym

A governed practice loop: take a real task, have an agent act, verify the result independently and keep a graded, private record of the attempt. It is built from components that already exist, not as a new training platform.

Research direction

Have a question
worth working on?

Talk to Odin