Skip to content
View patkusch's full-sized avatar
  • London
  • 21:25 (UTC +01:00)

Block or report patkusch

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
patkusch/README.md

I build the checks that stop AI agents doing what a bank would not allow, and I publish the experiments that failed as readily as the ones that worked.

Most of what I build sits underneath AI agents: keeping their work safe when they crash, checking that the notes they rely on are still true, and setting limits on what they may do without asking a person first. I spent years delivering change inside banks, which is where I learned that the failures that hurt are the quiet ones.

Everything here is public, free to reuse (MIT unless the repo says otherwise), and runs from a fresh download.


Selected work

What it is Why it matters
apiary Keeps a team of AI agents working when one of them crashes A worker that died used to take its job with it. Now the job goes to another worker, is retried a limited number of times, and is parked for a person if it keeps failing.
acta A record of what an agent did that it cannot rewrite A logbook the driver can write in but cannot tear pages out of, with an honest list of which tricks it catches and which it cannot.
cresco A radar for which tech skills employers will want next Every prediction is written down with a date and later checked against what actually happened, misses published alongside hits. Six years of real job adverts behind it.
kontext Tells you which of your project notes are out of date Compares each note with the code it describes and when each last changed, so nobody trusts a note that stopped being true.
remit Rules for how much an AI agent may do on its own Asks how much damage one action could do before it is allowed unattended, mapped onto the rulebooks banks and EU regulators actually use.
aurora Finds where two approved documents contradict each other Two design documents, both signed off weeks apart, cannot both be true. It finds the clash in seconds and points at the exact lines.
hazlog The same, for hospital IT safety Catches contradictions that could harm a patient, and keeps patient data on site: only single words ever leave for the cloud model.

Things I got wrong, in public

This is the part I'd actually point a reviewer at. Anyone can publish the run that worked.

  • A finding that died as the data grew. I thought Wikipedia page views predicted hiring six months ahead, and the numbers looked convincing enough to go in the README. Then I fixed a bug that had been cutting off the job-advert history, added more years of data, and watched the effect vanish. The retraction is in the README with the table showing it disappear, because a project about grading its own predictions doesn't get to quietly delete one.

  • Measuring whether my own instrument works. I asked four people, separately and without the answers, to classify eight agent failures using my manual. They agreed with each other far more than chance would allow (κ = 0.83, where 0.61 counts as good). They also found two mistakes in the manual. Both are fixed, and both are written up.

  • Separating what I wrote from what I inherited. apiary started as a copy of someone else's project. Its test suite is 3,763 tests and green; 3,691 of them came with the copy and I did not write them. The badges say so, with the file and line counts to back it up.


Background

Management consultant. I help banks and other regulated companies change how they work: working out what the business needs, delivering the product, running large programmes. Knowing the rulebooks (DORA, the EU AI Act, NIST's AI risk framework, ISO 42001) is why the governance work above is grounded rather than theoretical.

Pinned Loading

  1. reglens reglens Public

    A regulatory-change investment-decision agent built on DataHub — discovers impact via lineage, costs ACT NOW / DEFER / MIN. COMPLIANCE, and writes the decision back into the graph.

    Python

  2. regproof regproof Public

    Regulatory-change remediation copilot for financial services — the agent maps, groups and drafts; a human approves every cohort; every action is proven and audited. DORA demo.

    TypeScript

  3. ryo ryo Public

    Ryo — agentic AI café operator. Your Monday, already handled.

    TypeScript

  4. remit remit Public

    An agentic skills framework for AI governance. Autonomy tiers, blast radius, and a diagnostic manual for agent failure modes — over EU AI Act, NIST AI RMF, ISO 42001 and DORA.

    Python