Available for engagements
I build systems that can prove what they did.
Agents are taking real actions — moving money, changing records, contacting customers. The question that follows every one of them is the same: who did this, were they allowed to, what exactly did they do, and can you prove none of it has been altered since? I build the layer that answers it, and I have built the whole of one alone.
What I can do for you
AI governance audit
I instrument your existing agent deployment and deliver a signed record of what your agents actually did over an agreed window. Most audits end in a PDF of opinions. This one ends in a hash-chained artifact you keep, re-verify yourself with a public key, and hand to your own auditor without trusting either of us.
Deployment and integration
Getting the suite running against your stack — your models, your providers, your compliance boundary — including providers no vendor list names yet. Local and self-hosted first: Ollama, vLLM and llama.cpp, with 21 provider integrations in the registry.
The compliance narrative
Mapping what you already do onto the EU AI Act (Articles 12, 13, 14, 19, 26 and 72), ISO/IEC 42001, SOC 2 and the NIST AI RMF — including, in writing, where the evidence does not reach.
The work, in numbers I did not type
Why the defect log is the reference
ABSuite carries 51 numbered sections of its own defects, in public, in the repository — including the ones I diagnosed wrongly first, and the fixes that were themselves wrong. One entry records three consecutive wrong diagnoses and what finally ended them: making the failure report what it observed instead of what it assumed.
That file is the reference I would rather be judged on than any CV. Code shows what someone can build. A defect log shows what they do when they are wrong — which is the part you are actually buying, because on a long engagement it is the part that happens most.
The same discipline is enforced mechanically: 29 build gates, several of which exist because something shipped broken once and no test would have caught it. Two of the eight architectural layers are marked unbuilt and claim nothing, and a check fails the build if the interface ever claims otherwise.
Start a conversation
Tell me what your agents do and what you need to be able to prove about them. If ABSuite is the wrong tool for it, I will say so — that answer costs you one email and saves you a quarter.
Reach me through GitHub →