From Patent Filings to a Ranked Inventor Shortlist

A technical sourcer needs the engineers and scientists credited on patents in a specific technology area, turned into a reachable, ranked hiring shortlist. This playbook is for in-house sourcers, talent leaders, and founders hiring deep-tech talent who want to work only from public patent data. It runs the job end to end: from a CPC class or a competitor's portfolio to a deduplicated, ranked list of inventors, each resolved to a current employer and a reachability signal.

Most published advice stops at "search the inventor's name." That advice is broken by three traps it never mentions: name disambiguation, the gap between the assignee printed on the patent and where the inventor actually works now, and the filing-to-publication lag that makes every patent a lagging indicator. This guide is built around those traps. Each one gets a step and a failure mode, so you finish with a shortlist you can defend rather than a pile of names.

What a patent shortlist can and cannot tell you

A patent shortlist is a list of people who provably contributed to an invention in a technology area, ranked by a composite of credible value signals, and resolved to where they work today. It is a strong source of hard-to-fake technical credibility and a poor source of current employment, and you must treat those two properties separately.

The credibility is real. To be named on a patent, an inventor must have contributed to the conception of at least one claim; if there are several claims, contributing to one is enough. That is a lower bar than "led the project," but it is a hard, legal bar that cannot be padded with a title. The employment, by contrast, is stale by construction, and the rest of this playbook exists to work around that.

Three structural facts shape everything downstream. First, US non-provisional applications publish about 18 months after the earliest priority date under 35 U.S.C. 122(b), and the clock runs from the provisional date, not the later non-provisional filing. Second, average total pendency to grant runs about 22.8 months. Third, US-only filings with non-publication requests never publish at all. Together these mean the most recent 18-plus months of a target's R&D is invisible, and some inventors are structurally absent. Your shortlist is a map of where strong people were, not where they are.

The data source and its limits

The primary free source is the PatentsView Search API, at base URL https://search.patentsview.org/api/v1. It filters patents by CPC class and by disambiguated assignee organization, returns disambiguated inventors with persistent IDs, and exposes endpoints for patents, inventors, assignees, locations, and CPC classes. Every request needs an X-Api-Key header.

Two operational limits govern how you pull. The API allows 45 requests per minute per key; exceeding it returns HTTP 429 with a Retry-After header. Result windows are also capped, with some clients limiting to 1,000 records per query and others paging up to 10,000 per page against a 100,000-record ceiling. The database updates weekly, so a pull is reproducible within a week but will drift over time.

The reason to prefer PatentsView over a plain Google Patents search is disambiguation. PatentsView assigns a persistent inventor ID produced by a clustering method (Monath and McCallum, 2015) and a Jaro-Winkler model for assignees, and it publishes precision, recall, and F1 for each update. Those numbers are the single most important thing to internalize before you trust a count.

Read this table as a warning, not reassurance. Inventor precision of 0.863 means false merges exist in roughly one of every eight to nine merged records, and because precision sits below recall, the errors skew toward lumping two people into one ID rather than splitting one person. In practice that means your most "prolific" inventor is also your most likely artifact. Assignee disambiguation is cleaner at 0.949 F1, which matters because you filter on assignee organization to scope a competitor's portfolio.

Run the procedure end to end

The method has eight stages. Expect a focused day of work for a single technology area, split roughly as scoping and pulling in the morning, dedupe and ranking around midday, and employer and reachability resolution filling the afternoon. The time-consuming part is not the query; it is resolving current employers, which the API cannot do for you.

Scoping notes that save a rerun

Scope on CPC class, not keyword. CPC is the classification examiners apply, so filtering by, for example, G06N for neural networks returns the right inventions regardless of how the applicant titled them. Collect assignee variants before you query: a single company often files under several legal entity names, subsidiaries, and acquired brands, and because assignee disambiguation is good but not perfect at 0.949 F1, a variant you miss is a block of inventors you never see. Pick a date window deliberately. Given the 18-month publication lag, a window ending "today" is mostly empty at the front; a window of the last five to seven priority years captures a settled body of work.

Pulling without truncating

The quiet killer in step two is rate-limit truncation. If you blow past 45 requests per minute, the pull stops with a 429 and you are left with a partial, often date-skewed sample that still looks complete. Always capture the API's reported total count and compare it to the rows you actually retrieved before you move on. If the two disagree, you are ranking a biased fragment.

Score inventors so the rank survives scrutiny

Rank on a composite, never on a single field, because every individual signal lies in a predictable way. The two tempting shortcuts, inventor sequence order and forward-citation count, are exactly the two that fail, so a defensible rank blends several weak signals into one that is harder to fool.

Sequence order carries no legal weight and no relationship to contribution. The convention of putting the primary contributor first is unreliable and firm-specific, so ranking by first-named inventor assumes a primacy that does not exist. Forward citations are the most widely used value proxy and correlate with patent value across many studies, but they are noisy: citations are frequently added by the examiner rather than by a later inventor building on the work, so a high count can reflect examiner behaviour or an incremental follow-on rather than quality.

Combine the following, and expose each as its own column so a hiring manager can see the reasoning:

• Patent count in the target CPC - proves sustained focus in the area; lies when disambiguation has lumped two people, inflating one ID.
• Forward citations - proves downstream influence; lies when citations are examiner-added or reward an incremental patent.
• Claim counts - proves substance of contribution per filing; harder to pad than raw patent count.
• First or sole inventorship frequency - proves repeated central involvement across filings, which is more credible than a single first-named position.
• Repeat co-inventorship - proves a stable working relationship and a real research line, and helps confirm the disambiguation ID is one coherent person.
• Recency - proves the work is still relevant and the person was recently active in the area.

No single patent field proves primacy, so rank on a composite or accept that your top names are noise.

The reason to combine rather than pick is that the errors are uncorrelated. A lumping error inflates patent count but scatters CPC classes incoherently; examiner citations inflate citations but not claim counts; sequence games the first-named field but not repeat co-inventorship. A candidate who scores high across all of them is far more likely to be a genuine primary contributor than one who tops any single column.

Resolve the assignee to a current employer

This is the step the old how-tos skip and the step that makes the shortlist usable. There is no published canonical method for mapping a disambiguated inventor to a current employer and reachable contact; no regulator or platform documents one. The practical workflow is to take the disambiguated inventor plus assignee plus location and cross-reference an external professional profile and the open web to find the current role, treating the patent assignee as historical throughout.

Why this cannot be skipped: publication lag plus pendency means the assignee data is roughly two years old. The person credited on a Siemens or Google patent may well have moved, and moving is exactly the signal a sourcer wants. The example queries that make this work are explicit about it, asking for inventors "now working somewhere else" or who "left Meta or Google in the last two years and are now at an early-stage startup." The move is the opportunity; the patent is only the proof of skill.

This is also the step where describing the person precisely in plain language removes the most friction. Instead of scraping and matching by hand, you can state the composite you just built as a single request.

Flag confidence on every resolution. A name, a location, and an employer that line up cleanly with a public profile is high confidence; a common name with no corroborating location is low, and low-confidence rows should not go to a hiring manager as facts. Refolk is where I close this gap, because asking for the current state of a person in English avoids the manual cross-referencing that eats the afternoon.

Attach a reachability signal

A ranked name with no channel is not actionable. For each top candidate, record at least one of a public professional profile, a GitHub account, or a published email on a recent paper, and flag anyone with none as "no signal" rather than quietly dropping them. Researchers in particular often publish contactable emails on papers, which is why the battery-materials example query asks for "a reachable email on a recent paper." A no-signal flag is honest information; a missing row is a hole the hiring manager will find later.

How this goes wrong

The failure modes below are the difference between a shortlist and a liability. Each has a tell and a check, and the most dangerous ones cluster at the top of your ranking, which is exactly where a hiring manager looks first.

Lumping deserves the hardest scrutiny because the math predicts it. With precision at 0.863 and recall at 0.922, false merges are more common than false splits, and a merge inflates exactly the patent-count signal that lifts someone to the top of your list. When your number-one inventor spans wildly different CPC areas or co-invents with two disconnected clusters of people, suspect that the ID is two people, not one genius.

Splitting is the quieter loss. A strong inventor who changed name or moved cities can fracture into two IDs, each with half the patents, so neither clears your threshold and you never see the person. The only fix is manual: for names near your cutoff, search variants and location histories to see whether two low-ranked IDs are one high-ranked person.

The non-publication blind spot is structural and unfixable from patent data alone. US-only filings with non-publication requests and anything under a secrecy order never publish, so some of the best people in a sensitive area are simply absent. Treat the patent shortlist as one input and triangulate with conference papers and public GitHub for the same technology area, which surface people the patent record misses.

Scarcity: size the pool before you promise a shortlist

Before you commit to a patent search, sanity-check how many reachable people the technology area and geography can actually yield. A thin pool is not a sourcing failure; it is a fact to set expectations around, and the supply differences inside "AI" alone are larger than most hiring managers expect.

In Refolk's index, the specialism gap is steep. US computer-vision research scientists outnumber their NLP counterparts about 3.9 to 1, so an NLP-scoped patent search should expect a far thinner reachable pool and longer outreach cycles than a vision-scoped one, even if both start from a healthy set of filings.

Geography multiplies the same effect. The same ML research-scientist profile is about 11 times more common in the US than in Germany, so a Germany-scoped patent search converts a thin filing set into an even thinner hire-able set. If your target assignee is a European lab, plan for a smaller shortlist and a slower cycle from the start.

A reusable score template and a final check

Lock your scoring into a written rubric so every run is comparable and every rank is defensible. Below is a skeleton you can paste into a sheet; the weights are a starting point, not a law, and you should tune them to the technology area.

score = 0.30 * patent_count_in_cpc
      + 0.20 * forward_citations
      + 0.20 * claim_count_total
      + 0.15 * first_or_sole_inventor_rate
      + 0.10 * repeat_co_inventor_rate
      + 0.05 * recency
columns to keep visible: inventor_id, name, patent_count, citations, claims, first_sole_rate, co_inventor_rate, last_filing_year, current_employer, employer_confidence, reachability_channel
flags: lumping_suspected (y/n), split_suspected (y/n), no_signal (y/n)

Before you hand anything over, run the shortlist against a closing checklist. The goal is that every claim in the file is one you could defend to a skeptical hiring manager.

Keep the shortlist current

A patent shortlist decays, so treat it as a living artifact rather than a one-time deliverable. The PatentsView database updates weekly and new inventions keep publishing on the 18-month lag, so re-pull the CPC and assignee scope on a regular cadence and diff the new inventor IDs against your last run to catch people who have just crossed the publication threshold.

The faster-moving half is employment. Because the assignee is roughly two years stale the moment it publishes, the current-employer resolution is the field most likely to be wrong next quarter, especially for the people you most want: those in motion. Re-resolve current employers on any candidate you did not reach the first time, watch for the moves that turn a patent into an opening, and keep the reachability channels fresh so that when the hiring manager says go, the shortlist is ready rather than two years out of date.