SOTERIA
An AI platform white-labelled for an intellectual property expert firm. It runs prior-art search for patents and trademarks directly in the official databases (EPO, USPTO, INPI, EUIPO), answers IP law questions, assists patent drafting and compares documents. Self-hosted open-source LLMs: no client data ever leaves for a third party.
Private demo
This client project is confidential. The screenshots above are shared with client approval.
// recruiter view
AI platform white-labelled for an intellectual property expert firm: prior-art search for patents and trademarks in the official databases, RAG chatbot over IP law, patent drafting and document comparison assistants. Carried from architecture to a live SaaS monetised through subscriptions — on self-hosted open-source LLMs, a non-negotiable requirement on confidential IP files.
- ▸Delivered white-label: the client firm operates the platform under its own brand
- ▸Prior-art search for patents and trademarks directly in the official EPO, USPTO, INPI and EUIPO databases — the most time-consuming work in the trade
- ▸RAG chatbot over intellectual property law, with sourced answers
- ▸Patent drafting and document comparison assistants
- ▸Self-hosted open-source LLMs (Qwen / Mistral): no client data on third parties — a decisive argument on confidential filings
- ▸From architecture to a live SaaS, monetised through Stripe subscriptions
The story behind
- Chapter 01
The need
An intellectual property expert firm wanted its own AI platform, under its own brand. The pain point was precise: before every patent or trademark filing, its experts spend hours searching for prior art database by database — EPO, USPTO, INPI, EUIPO — then reading and comparing documents. Expert time billed for search work.
- Chapter 02
The constraint that decided everything
An unfiled IP case is one of a company's most sensitive assets. Routing those documents through a third-party LLM API was a non-starter from the first conversation. The constraint became the architecture choice: open-source LLMs (Qwen / Mistral) self-hosted on Kubernetes at OVH, no client data on third parties.
- Chapter 03
Prior-art search
Four official databases, four formats, four query models. The core of the product was wiring them directly and normalising their results into a single model, so the expert sees one list of relevant prior art rather than four tabs to reconcile by hand.
- Chapter 04
Law, without hallucination
On IP law, an invented answer is worse than no answer. The chatbot therefore sits on a RAG pipeline over the legal corpus, with chunking tuned for statutes and patent claims, and systematically sourced answers — the expert can trace back to the source before committing.
- Chapter 05
From prototype to billed product
The platform did not stay a demo: React 19 frontend, FastAPI backend, PostgreSQL, containerized deployment on Kubernetes OVH, and Stripe subscriptions wired in for monetisation. Delivered white-label, it is operated by the firm under its own identity.
After going live
Monitoring & observability
Production observability on Kubernetes OVH: inference pod health, official-database connector latency, RAG quality and subscription tracking.
Stack
Tracked metrics
- ●Self-hosted LLM inference latency (Qwen / Mistral)
- ●EPO / USPTO / INPI / EUIPO connector latency and error rate
- ●Relevance of prior art surfaced per search
- ●Share of sourced answers from the RAG chatbot
- ●Active subscriptions and Stripe billing events
Production impact
Shipped to production and operated white-label by the client firm, the platform moves prior-art search work from the expert to the machine.
- ▸Prior-art search automated across all four official databases instead of manual database-by-database lookups
- ▸IP law questions handled by a RAG chatbot with sourced answers
- ▸Patent drafting and document comparison assisted
- ▸Confidentiality held end to end: no client data on third parties
- ▸Billed through subscriptions — not an internal prototype
// results
// stack