AI is a technology. That is why it changes so much.
After 20+ years of building IT systems, I treat AI like any other technology: it has to fit the processes, the data and the systems a company already runs. And because it is a technology and not magic, it can also do things no system could do before. I work on both sides.
AI inside classic product development
In a real system, AI is one component among many, next to databases, integrations, rules, user rights and logs. It reads a document, routes a request, drafts a reply, and everything around it stays proven engineering: clear interfaces, tests, releases and an owner. That is what makes it reliable enough for daily work.
AI doing what was not possible before
Some things from my own systems:
- An investigation without a single news article. I let an AI agent work out what happened in a crypto exchange hack using only the API of my own blockchain explorer, on-chain data and nothing else.
- A newsroom that checks itself. Every day a pipeline reads hundreds of news items, writes briefings in two languages and narrates them in my own cloned voice. Then it listens to its own episode and checks that it says what the text says.
- AI that builds the product, then steps aside. In one product, AI wrote thousands of texts once and I approved them. At run time no model is called at all: instant, almost free, the same answer every time.
- 30% faster on the same hardware. I combined two open-source techniques for running a 31B model and measured the gain, with numbers anyone can check.
The fundamentals do not change
- Data quality. Bad data in, confident nonsense out. Most AI problems start here.
- Documents and OCR. Much company knowledge still lives in PDFs, scans and email. Getting clean text out of them is half the job.
- RAG. Answers from your own documents, with the passage they come from. It is a search problem before it is an AI problem.
- Local models. Open-weight models on your own hardware, compressed to fit, when data must not leave the building.
- Integration. AI is only useful inside the tools people already use: mail, document stores, business systems.
- Measurement. A set of real cases with known answers, run again after every change of model or prompt.
What I build
- AI & Automation Audit. Where AI really pays off: the work worth automating, ranked by value, effort and risk, the right setup for each case, and its cost over three years. Plain automation first, where it is enough.
- Knowledge Assistant (RAG). An assistant that answers from internal documents, such as policies, contracts, product information and guides, and always shows the source.
- AI Governance. Rules people actually follow: the AI tools in use, an approved-tools list, a risk register, a short AI policy and training. Not legal advice; where legal sign-off is needed, I work with a lawyer.
- Fine-Tuning. When good prompts and document search are not enough: a fixed output format, fields taken from documents, the language of an industry, or a smaller model doing the job of a big one. Always measured on a test set, before and after.
- Keeping it running. After the build: document updates, answer quality, users and access, model updates and monitoring.
Where it runs
- Frontier models in the cloud for data that is not sensitive: the leading public models, with the right data settings and access rules.
- Private AI on Swiss servers when data must stay in the country: a private AI workspace in a Swiss data centre.
- Local AI in the office when it must never leave the building: a server on site with open-weight models.
From my own systems
- btc.artjoms.ai: 678 million Bitcoin addresses, plus USDT and USDC, checked against sanctions lists and issuer freezes.
- Signal: a daily briefing on AI, markets and regulation, in two languages, with video.
- Legal RAG: law answers with the exact article. Private, access on request.
- Quantised models: open-weight LLMs that run on your own machine, downloaded thousands of times.
Thinking about the same questions? Write to me through the form or on LinkedIn.
Questions or ideas?