The problem, before.
Patent drafting is one of the few places where precision of language has legal consequence. A misplaced claim, a vague disposition, a forgotten fallback — each can narrow or invalidate a right. Yet most drafting work is repetitive: transposing an inventor’s disclosure into the grammar of patent specification.
Before
An attorney reads a raw invention disclosure — prose, figures, sometimes code — then manually produces background, summary, detailed description, brief description of drawings, and a claim set. It is careful, repetitive, high-stakes work that dominates attorney time and gates client throughput.
After
PatFace ingests the same disclosure and returns a structured draft application — every section written in the expected voice, with prior-art neighbours cited, claims validated for antecedent-basis, and a consistency critic flagging any cross-section drift. The attorney refines; they do not rewrite.
The system had to preserve what is legal about the work while automating what is mechanical. Hallucination, here, is not a UX bug — it is a liability risk. Grounding was the first-order problem, not a polish.
“His understanding of both frontend and backend made our sprint cycles significantly faster. He does not just execute tasks, he improves the process around them.”
Harnessing agentic AI.
A single LLM call cannot hold the whole job in its head. Drafting is a series of distinct cognitive tasks — identify the invention, retrieve prior art, structure claims, write narrative, validate internal consistency. We assign each to a specialised agent, and let a coordinator choreograph them.
This is the agentic pattern at its most deliberate: agents are not wandering generalists but tightly-scoped specialists. The coordinator does not reason about patents; it reasons about workflow. The specialists do not reason about workflow; they reason about patents.
Building an agentic AI system for your domain?
Whether it's legal, medical, finance or operations — agentic systems work when the domain is encoded in the agents, not hand-waved into the prompts. 30-minute scoping call for clarity on where the real complexity lives.
RAG — grounding the agents.
Patent drafting without grounding is hallucination waiting to happen. The retrieval layer keeps the agents honest: it provides recent art, cited prosecution history and examiner rejections to neighbours of the claimed invention, before a single line of draft text is written.
Theory of strong design — for agentic systems.
An agent is strong not when it does more, but when the space of things it could plausibly do is small and legible.
Three principles held the design together. One: every agent has a schema, not prose, as its output — structured JSON the next agent can consume without parsing English. Two: the coordinator never writes prose. Its job is to route, retry and reject. Three: the critic is the last word. A consistency critic reads the assembled draft and can send any section back for revision, with concrete objections.
The result is a system that behaves like a small practice group: a partner planning, associates drafting, a senior reviewing — each stayed in their lane, each replaceable with a better model as the state of the art moves.
How it was built.
Build timeline
- Domain framingWeek 1–2
Schema for each draft section, rubric for what a filable draft looks like, examiner-rejection taxonomy.
- RAG corpusWeek 3–4
Ingestion pipeline for USPTO/EPO/WIPO passages. Hybrid search and re-ranker tuned against held-out queries.
- Agent scaffoldingWeek 5–6
Coordinator, specialists, critic — each with typed JSON contracts and token budgets.
- Rendering + guardrailsWeek 7–8
Claim dependency validator, antecedent-basis check, §101 flags, DOCX/PDF output.
- Calibration with attorneysWeek 9–10
Side-by-side drafts on real disclosures, prompt tuning, prior-art overlap calibration.
Outcome.
Live since 7 March 2026. 450+ patent applications drafted through the platform at a steady 10 applications per day. The system takes an invention disclosure and returns a draft application that a qualified attorney can refine, not rewrite. PatFace is where the author’s patent-domain expertise meets the author’s AI engineering — a rare intersection, and the reason the output feels like it was written by someone who has signed a specification before.
Have a domain where language is the product?
Legal, medical, compliance, research, insurance — domains where output correctness is non-negotiable are where agentic AI actually earns its keep. Book a 30-minute call and we'll map yours.