Ideas, stories, and creative sparks

Designing an AI That Actually Knows the Rules (Aug 2025)

‍

We’ve all seen chatbots that spit trivia and hallucinate with the confidence of a mediocre consultant. That’s fine for fun. It’s not fine when a wrong answer means misread policy, missed entitlements, or institutional risk.

That’s the problem I set out to solve.

I built a fully local AI trained only on the OSCE Staff Rules and Regulations. Not to show off, but to fix a daily pain: most staff don’t know where the rules live—and even when they find them, they’re unsure it’s the right section, version, or context.

Organizations bury knowledge in PDFs and SharePoint. Search doesn’t help. Asking around gets you ten opinions and one passive-aggressive email. The issue isn’t a lack of rules; it’s a lack of structured, accessible truth.

This chatbot changes that. It runs locally with Mistral and ChromaDB via llama.cpp—fast, compliant, secure. No external API calls. No guesswork. Every answer is sourced from the actual Staff Rules. It’s multilingual-ready. It looks like ChatGPT, but it actually understands the OSCE’s legal framework and shows its sources.

It’s more than a chatbot; it’s a new pipeline for institutional knowledge. It removes gatekeepers and gives staff, field missions, and auditors direct access to the exact clause they need.

Building it exposed a deeper truth: most internal systems aren’t designed to help people—they’re designed to avoid blame. They’re slow, vague, and wrapped in forms. This AI flips that. It assumes the user is intelligent, time-starved, and needs a correct decision now.

And it works. It explains complex clauses, cites sections, tells you what isn’t covered, and when to escalate to HR. It doesn’t bluff. It respects the user and shows its sources.

It’s not flashy. It’s useful. And it’s a blueprint for applying AI where accuracy and accountability actually matter.

This is just the beginning.

‍