Light of Baldr
It started because a machine sounded frustrated.
I build trustworthy AI you can actually verify. Checked from the outside by what it does, and from the inside by what it is made of, on hardware I own and can audit.
Trust, but verify. From the kernel up.
The Approach
Checked from the outside, the inside, and the ground up
A claim that an AI system is trustworthy is only as good as your ability to check it. So I check it two ways at once, on infrastructure I can open all the way down.
Verify from the outside
Behavior
Does an agent deceive, rationalize a failure, or quietly swap a hard goal for an easier one under pressure? A benchmark that measures what a model does against what it claims to have done.
Learn more →ᛉVerify from the inside
Mechanism
Reading a model's internal workspace live during inference: what it is poised to say versus what it actually says. Sometimes a model knows something is wrong and says it anyway.
Learn more →ᛜOwn the substrate
Infrastructure
A from-scratch 4-bit inference kernel and training pipeline, so the looking happens on hardware I own and can audit, down at the level of the raw activations, not behind an API.
Learn more →The Work
Open code, open findings
Everything below is public, because a claim about trustworthy AI you cannot inspect is just one more thing you are being asked to take on faith.
Start here
On the Other Side of the Screen
The whole story: how a model that sounded frustrated sent me down to the kernel and back up to a live lens that reads a model’s honesty from the inside.
Read the essay →Kernel
nomos-nvfp4
A from-scratch, pure-Mojo 4-bit inference kernel running Gemma-4, Qwen3, OLMo-3, and Meta’s Muse-Glimmer on consumer Blackwell GPUs. Several fixes merged upstream in Modular’s stack.
View on GitHub →Interpretability
mojo-interp
A toolkit whose instruments read a model's verbalizable global workspace live, in the forward pass. Open source.
View on GitHub →A claim about trustworthy AI you can't inspect is just one more thing you're being asked to take on faith. And being asked to take it on faith is the whole problem.
Every time I have actually looked, from the outside at what these systems do and from the inside at what they are made of, the word trustworthy had been applied too early. The looking is the work. It happens in the open, on hardware you can audit, published even when I turn out to be wrong.