2025 - The year of shipping advanced AI
02 Mar 2026Satisfying year of working on generative AI products at Wolters Kluwer, this time across three countries and four releases. Each project had its own team, its own challenges, and its own lessons. On top of everything, I enjoyed close cooperation with engineering teams and product teams.
Four AI releases
1. VitalLaw Expert AI — conversational search (USA). Building on the foundation from the previous year, the product expanded with a conversational AI feature that lets legal professionals query primary law and expert commentary through natural language. A big part of my involvement was building a proper evaluation layer — getting the right datasets and making sure improvements were technically measurable rather than just subjective.
2. ONE AI — generative AI for professionals (Italy). One business unit, but in practice something closer to three sometimes overlapping workstreams: legal, tax, and labour — partially similar, but with different data sources and evaluation datasets, optimisations and stakeholders. On top of that, AI UX features were being developed and integrated into the chat interface. A lot of coordination, but also a staying close to the timelines.
3. LEX Expert AI — specialised legal AI chat (Poland). Two beta phases, close cooperation with the business side, and a lot of focus on quality. The interesting problems here were around understanding user intent and getting the balance right between primary sources (legislation, rulings) and secondary ones (expert commentary). Working with a very diverse group engineers from multiple teams, proving commitments comes hand in hand with AI expertise for project success.
4. LEX Jurisprudential Compass — auto-summarization (Poland). My contribution was within the auto-summarization part of this tool, which helps lawyers make sense of a rapidly growing body of court rulings. Relatively smaller to other projects but with a scale potential as across two years also Germany, Czechia and Hungary benefited from similar framework.
Staying close to the technology
Personal project. I’ve been building a side project that combines machine learning and generative AI to create news digests from high-quality sources. It’s where I get to make all the calls myself and run experiments without a delivery deadline attached.
MCP. Completed the Model Context Protocol course from DeepLearning.AI — a useful deep-dive into one of the more practically interesting protocol developments in the AI tooling space.
Hackathon. Took part in an internal hackathon and found a real defect in the company’s AI platform in the process — one that got fixed within weeks. A good reminder that hands-on engagement with tools, even briefly, surfaces things that reading documentation never would.
Claude Sonnet 4.5 is a beast Started the year with GPT being a solid companion for writing isolated code and reading logs for me. When Claude Sonnet 4.5 was released, it was a game-changer. Generating fully working Lambda functions is a single shot, turning experience-based prompts into fully working code was astonishing. It was “GPT-3 wow effect” for me.
Reading, listening, learning
The Lean Startup by Eric Ries. Still relevant, maybe more so in the context of AI products where it’s tempting to over-build before validating. The build-measure-learn loop is simple advice that’s easy to ignore under deadline pressure.
The Pragmatic Engineer by Gergely Orosz remains my most-recommended read in this space. It covers AI’s effect on the industry, engineering culture, and the job market in a way that feels grounded rather than hype-driven. Good for staying calibrated.
Lenny’s Podcast — particularly the episodes on Shape Up. The framing of appetite over estimates and the separation between shaping and building gave me useful language for conversations I was already having. Not a methodology I’d apply wholesale, but the thinking behind it is worth sitting with.
A few observations
Evaluation is king. The gap between “looks good in a demo” and “works reliably for users” is where AI projects quietly fail. Securing proper datasets and defining honest metrics is slow, unglamorous work — and exactly the kind of thing that gets skipped when delivery pressure builds up.
Clarity on direction enables agility, not the opposite. Fuzzy requirements and leaving key processes undefined don’t create room to pivot; they remove the stable ground you need to pivot from. Shape Up’s core insight applies: fix the appetite and the problem worth solving, then give the team room to figure out how.
Translation is a real skill. Getting the same message to land for an engineering team and a business stakeholder requires more than just simplifying. It’s about finding the frame that makes the tradeoff legible to each audience without losing the actual content.
Thanks for reading. Let’s stay in touch.