ClaimForge

Patent monetization was bottlenecked on manual claim charting. A patent licensing team held a portfolio worth pursuing, but every claim chart — the document mapping a patent's claim language to evidence in a specific product — took a patent attorney days to build by hand: reading claim elements, searching product documentation and teardown reports, and manually scoring how well each piece of evidence matched. With a portfolio of hundreds of patents and thousands of potential infringement targets, most of the portfolio's value sat unexamined simply because there weren't enough hours to look.

The team didn't need another patent search tool. They needed the claim-to-evidence mapping done for them, with a defensible score and citation trail an attorney could actually stand behind.

A pipeline from claim language to scored evidence. ClaimForge ingests a patent's claims and a target product's documentation (spec sheets, teardown reports, technical manuals) and produces an element-by-element mapping: each claim element matched against the strongest supporting evidence, with a confidence score and the source passage attached. I built the system end to end: the claim-parsing pipeline, the evidence-retrieval and scoring engine, the attorney-facing review app, and the portfolio-level analytics layer.

Claims are broken into individual elements, evidence is retrieved from a searchable corpus of product documentation, and each element-evidence pair is scored for how well the language actually supports infringement. Attorneys review a decision-ready chart instead of building one from scratch, and every score traces back to the exact source sentence it came from.

Score with evidence, never with a bare number. The core design decision was to make every score explainable. A parsing service breaks each claim into discrete elements. A retrieval service searches the target product's documentation for passages relevant to each element. A scoring model then rates how strongly each passage supports that specific claim element, and returns the score alongside the exact source text, never a number on its own.

This kept the tool usable in a legal context, where an unexplained AI score is worthless. Attorneys could see exactly which sentence drove a 91 versus a 45, correct a mismatch in one click, and trust the chart enough to build a licensing argument on top of it.

Built for a field where a wrong claim chart has real consequences. Because claim charts underpin licensing negotiations and litigation, I treated citation accuracy as the core reliability metric, not overall speed. Every score is logged with its source passage and model version, so a chart could be re-verified months later if a negotiation revisited it. Low-confidence matches were flagged for manual review rather than silently included, and the retrieval index was rebuilt on a schedule to keep pace with newly indexed product documentation.

Attorneys were cautious about trusting an AI-generated chart until the citations held up under their own scrutiny, at which point claim charting stopped being the bottleneck on which parts of the portfolio got pursued at all.

Claim charting stopped being the constraint on which parts of the portfolio got pursued at all — attorneys review a scored chart instead of building one from scratch.