Thedy & Partners.

The problem

Preparing a tax opinion took hours of manually searching and analysing interpretations. The knowledge was scattered across a corpus of more than half a million documents — with new ones arriving daily.

What I built

A system that maintains its own knowledge base and serves it to advisors in seconds:

• Half a million documents processed — to start, every one of the 538,000+ documents had to be analysed by AI: summarised, described with metadata and indexed. I spread an operation of that scale in parallel across several model providers' infrastructure; today the system runs on Anthropic models.
• A self-updating base — every night the system checks the Ministry of Finance's new publications and automatically downloads, analyses and indexes anything it doesn't yet have — retrying failures by itself, with no one on call.
• Hybrid search — an advisor's query is searched semantically and full-text at the same time; results are merged, scored for relevance and ordered before they reach the advisor.
• Opinion drafts — the system analyses the retrieved interpretations and prepares a preliminary opinion in two steps: analysis, then synthesis.
• A research agent — an autonomous AI agent that decides for itself when to query the search engine and when to pull a document's full text — conducting research in stages, the way a person would.
• The firm's portal backend — search, research and the agent are available to advisors through an internal portal.

Stack: n8n, PostgreSQL, Elasticsearch, Next.js; Anthropic models.

The outcome

Preliminary opinion preparation cut from around 12 hours to around 2 minutes. The engagement is ongoing: the system is being extended with further corpora — a court-rulings database and a legal-provisions database — on the same proven architecture.