AI systems that run unsupervised in front of real customers.

The firm was drowning in repetitive intake. I built an agent with 11 tools: four that retrieve from their legal corpus, three that act (send a document, escalate, book a call), four that manage its own internal state.

The interesting work wasn't the prompt. It was deciding what the agent isn't allowed to do. It never does arithmetic, because it quotes fees, so it calls a whitelisted evaluator instead. It doesn't guess: without country, qualification and year, it asks rather than answering a legal question. A four-level lead temperature decides whether a human gets interrupted at all.

The owner types an order into a chat. The system issues the invoice, records the expense, updates the CRM and feeds a dashboard. It replaced the management software they were paying for.

22 tools in the agent, a separate read-only role for their accountant, and EU VAT validated against the official registry rather than a regex.

A full learning platform: courses with content released on a schedule, progress tracking, billing and role-based access, built and shipped for a client with real students on it.

The community wall went from brief to production in about ten days. Real-time posting, moderation and access rules tied to what each student had actually paid for — the interesting part was never the feed, it was making sure the wall and the billing agreed on who could see what.

No menus, no forms, no flow to follow. The user just talks, and keeps talking — for months.

A conversation that long outgrows any context window, so memory is tiered: Redis holds the live conversation, Postgres holds what has to survive it. Answers are retrieved from the corpus rather than packed into the prompt, which is what keeps the cost flat as the history grows.

The memory of a conversational agent isn't a database problem. It's an editorial one: deciding what deserves to be remembered.

A CRM and a training platform in one product, running for three organisations across financial advice, insurance, banking and legal. Their consultants train inside it — video, documents, resources — and work the pipeline in the same place: lead through to client, then post-sale tracking built so cross-selling has somewhere to start from. It also carries the presentation deck a consultant walks a client through, writing their financial snapshot straight into the CRM as the meeting happens.

It isn't three products. It's one template, cloned three times — which is an advantage right up until a clone arrives quietly missing fourteen columns and the client tells you the numbers aren't saving.

Then there's language: three locales, one of them Greek, around 3,600 keys. The untranslated-string detector reported zero while half the deck was still in Spanish, because it was validating the dictionary and not the code. The fix wasn't reviewing harder. It was a gate — a lint rule that errors on literal strings, and a parity check that blocks the commit when the three locales drift apart.

An AI assistant sits on top with two jobs: operating the platform, and coaching the consultant before a client meeting. Built once, shipped three times. That's the point, not a caveat.