Andi Albrecht
Software architect · Python since 2002 · applied AII build things that have to hold.
Two decades of hands-on architecture. These days I take AI from idea into production in business software, where a decision still has to be explainable long after it was made.
NowJuly 2026
AI that has to survive an audit.
For the past two years I’ve been bringing LLM features into a grant-management platform: a domain built on dense rules, where an assessment has to be traceable and defensible months later, to someone who wasn’t in the room.
Assessments
Live in production: the model checks an application against the rules that apply to it and has to show its reasoning. The caseworker decides, with that reasoning in front of them. Human in the loop is a property of the architecture here, not a sentence in the terms of use.
Agentic architecture
The assistant earns its abilities from the system it serves, so there is no second integration layer for somebody to keep in sync by hand. It inherits the permissions already in place; one that can see more than the person using it is a data leak with good manners. Retrieval is a building block, not an architecture.
Operating it
Nothing is really in production until you can see what it does, so: tracing, and evaluation runs you can repeat. Prompts are product content. The people who own the wording edit it themselves, inside guardrails they cannot remove by accident.
Where I stop. No model training, no foundation-model internals, no data science. Embeddings, context windows, tool calling, eval strategy: those I use with confidence. How exactly a token becomes a vector is someone else’s field, and they are better at it than I would be. My strength sits one layer up, in the application and in keeping the thing running responsibly. I would rather say that than blur it.
Track record
Work that outlived the project it came from.
Open source is where engineering happens in public, by consensus, without a manager to break the tie. Some of mine is still running fifteen years later.
Making
The same impulse, after hours.
At thirteen I was drawing geometric figures on a CBM, one BASIC line at a time, with about as many pixels as patience. Today you would call that algorithmic art. I never really stopped: the field at the top of this page is the same idea, one rule on a fixed grid.
Approach
Judgement, not a rulebook.
The goal is always sustainable software: it works, it’s pleasant to use, and it stays maintainable, extensible and fast enough. That is five or six balls that partly contradict each other. The job is weighing them deliberately, and knowing which cost you just accepted.
I came to software from the humanities, with a background in philosophy. It taught me to frame a problem instead of forcing it, and to stay with a strange, complicated domain until it comes clear. That is the core of both architecture and applied AI, and I use it most weeks.
I read the Zen of Python the way Brian Eno meant Oblique Strategies: as a provocation to think, not as law.
Contact
Reachable, and interested in hard domains.
If you’re putting AI into a process where being wrong has consequences, or you need someone who can hold an architecture and still write the awkward part himself, say hello. I read everything that arrives, and I answer.