Agentic AI for Patents
and Tech

Leif supercharges your ability to make informed legal decisions by fully contextualizing your matter against all filed patents, scientific documents, technical information, and relevant web pages.

Patents and Tech are your company’s most important assets. Leif maps ownership, conflicts, and white-space across your portfolio — so you can see where you are protected and where you are exposed.

IP risk is hard to assess in a deal window. Leif assesses patent strength, ownership, threats, and white-space — and returns a diligence-ready report on risk and commercialization potential.

Leif brings best practices to client work for investment banks and consulting firms — connecting patent intelligence to ownership, threats, and opportunities.

Overview

LeifAI produces three proprietary scores. Each is a single number from 0 to 100; a higher value always indicates stronger support for the proposition the score names.

Every score is computed from stored evidence, per-element judgments, and a fixed, versioned system of scoring rules, so any result can be audited and reproduced from its underlying record.

What this system is, and is not. The scores are not the output of a single prompt to a general-purpose language model. They are produced by a purpose-built pipeline: corpus-scale dense-vector retrieval over tens of millions of embedded passages, multi-channel recall merged by reciprocal-rank fusion, instruction-conditioned cross-encoder ranking on dedicated GPU infrastructure, graph queries over ownership, legal status, and prosecution history, multi-sample consensus voting, role-based element weighting, and layered arithmetic aggregation. A language model participates at one bounded stage, rendering element-by-element chart verdicts with verbatim evidence. It does not select the references, rank them, assign the weights, or compute the score.

The Shared Scoring Engine

All three scores run on one pipeline of five stages; each score differs only in what is compared against what and in how the final number is oriented.

The evidence corpus. Every score is computed over a defined evidence corpus rather than a model's memory: more than seven million United States patents and published applications from official USPTO bulk data, together with their prosecution, ownership, and legal-status records, enabling the scores to distinguish live, enforceable claims from expired or pending ones. For retrieval, the corpus is embedded as nearly nine million whole-document vectors and more than fifty million passage vectors from a high-dimensional transformer bi-encoder that encodes indexed text and queries asymmetrically, a retrieval-specific design that measurably improves precision. The full text of every document is retained and is the source of every verbatim evidence quote a score cites. Filing and priority dates are indexed with every vector, so statutory prior-art cutoffs are enforced within the search itself.

Stage 1 — Decompose into elements. A claim or disclosure is broken into its elements, the discrete limitations that must each be satisfied under patent law. An independent claim splits into a preamble and an ordered list of limitations; a disclosure splits into its technical features. Each element receives a stable identifier and is tracked through the rest of the pipeline. Where an automated split would be degenerate, a verbatim re-split forces each element to appear as an in-order substring of the source text, a guard against fabricated limitations.

Stage 2 — Retrieve the most relevant prior art. The technology and each of its elements are encoded and matched against the corpus in a funnel that moves from broad recall to narrow precision, so that the system reads a large candidate pool while reserving its most expensive analysis for the most consequential references. Recall is multi-channel by design; no single method finds all relevant art.

• Document lane — whole-document semantic search for the globally closest patents.
• Element lane — per-element passage search for the most on-point text under each limitation.
• Second-model lane — a patent-claim embedding model, so that art missed in one vector geometry is recovered in another.
• Lexical lane — full-text keyword search for art whose relevance turns on exact terminology.
• Expansion lane — controlled query expansion, restating each element in alternative technical vocabulary.
• Document-as-query lane — search with the specification itself, bridging the vocabulary gap between older art and modern claim language.

Reciprocal-rank fusion merges the lanes, favoring references that rank well across several of them. A cross-encoder then re-reads each survivor jointly with the query at the token level, a more precise and more expensive comparison than vector similarity, and keeps only the strongest. It is instruction-conditioned to the legal question at hand, element-level disclosure versus infringement relevance, and runs on dedicated GPU infrastructure.

Stage 3 — Judge each element for consensus. For each surviving reference the system produces an element-by-element claim chart. A low-temperature language-model judge renders one of three verdicts per element: present (the reference or product satisfies it), unclear (the text is silent or ambiguous), or absent (affirmatively different or missing), each with a verbatim evidence quote, a short rationale, and a confidence value. Each judgment is sampled several times independently; the majority position becomes the verdict, and the reported confidence combines how strongly the samples agreed with how confident they were. A fast first-pass screen promotes only the most promising references to the full multi-sample vote.

Stage 4 — Weight elements by enforcement importance. Not every element carries equal weight. Each element is assigned a role, and each role a fixed weight, along an enforcement-importance ladder; the numeric weights are proprietary calibration values. The supporting and inventive roles are assigned by a dedicated inventive-aspect analysis that identifies the claim's true point of novelty from prosecution history and office-action evidence rather than surface wording. The weighting is decisive in clearance and in the point-of-novelty theory of validity, where the inventive element carries several times the influence of preamble language. The anticipation analysis, by contrast, weights every element equally, as the law requires a single reference to disclose them all.

• Preamble — lowest emphasis.
• Ordinary limitation — baseline emphasis.
• Supporting element (point-of-novelty cluster) — elevated emphasis.
• Inventive element (the point of novelty) — highest emphasis.

Stage 5 — Aggregate deterministically in layers. The reduction from judged evidence to a number is deterministic but not a single formula: it is a layered aggregation, each layer encoding a principle of patent law, applied in fixed order.

• Match strength. Each verdict becomes a continuous match strength scaled by its consensus confidence; ambiguous verdicts earn only a small fixed partial credit.
• Role-weighted coverage. Match strengths are combined under the Stage 4 ladder and normalized, so coverage reflects which elements matter, not how many are met.
• Statutory guardrails. Ceilings and escalations impose the law's structure on the arithmetic: full coverage escalates regardless of the average, and scores are capped progressively as unmet elements accumulate, because a theory that leaves elements uncovered cannot anticipate.
• Theory-level modifiers. Combination theories are discounted by the credibility of the reason to combine and by combination size; point-of-novelty theories are suppressed entirely unless a reference teaches the inventive element.
• Headline composition. Surviving theory scores are composed into the headline, oriented to the score's meaning, banded into tiers, and subjected to the indeterminacy rail.

The functional forms and constants inside each layer are proprietary. What is published is the architecture and a guarantee: every layer is fixed arithmetic over stored evidence, no step after the element verdicts involves a language model, and re-running the aggregation over the same record reproduces the same number.

Determinism, Validation, and Honesty Rails

Three properties make the scores defensible to attorneys and decision-makers.

Deterministic, not stochastic. Every score is produced by a fixed, versioned system of rules applied to stored evidence and judgments. The judgments are generated by the proprietary retrieval stack and low-temperature judges and are cached under content-addressed keys, so the same input reproduces the same output. The same discipline applies to infrastructure failures: if the ranking service is unreachable, the run fails visibly rather than proceeding on fabricated similarity scores.

Element by element, and weighted by what matters. A score is never a holistic guess. It is built from per-element verdicts, each backed by a verbatim evidence quote and weighted so that the true point of novelty outweighs boilerplate language, with risk banding that reflects the all-elements rule.

Honest about its limits. When the analysis cannot complete reliably, for example when too large a share of the reference comparisons fail, the system withholds the score and reports the result as indeterminate rather than publishing a misleading number. The validity pipeline is benchmarked against real Board decisions. All scores are bounded by the corpus, centered on modern-era United States patents and published applications and (for now) excluding non-patent literature and foreign art outside that set; each report states this boundary explicitly, so that a result is never mistaken for an exhaustive search.

Together, these properties reduce three difficult legal questions — freedom to operate, patentability, and validity — to three reproducible and defensible numbers.

Leif Clearance Score (LCS), Leif Patentability Score (LPS), and Leif Validity Score (LVS) are proprietary scoring methods of Leif Legal AI. The outputs are decision-support analytics developed over a defined United States prior-art corpus and do not constitute legal advice.

LeifAI was founded in 2026 and is based just outside of Washington D.C. We provide domain-specific AI solutions for patents and tech to attorneys, investors, companies, and advisors.

Leif three core values guide our product, business, and growth - customer-always, true excellence, and most useful.

Customer-Always — As founders who have worked in senior leadership roles in major companies, we know that true commitment to customers is the foundation of everything.

True Excellence — We founded the company to be the best at one thing. Stay focused to deliver excellence. Best product. Best service. Keep working to achieve it and maintain it.

Most Useful — The technology we use is complex, and the subject matter of patents and intellectual property is intensely detailed. We strive to make our product easy to use. We aim to take the complicated and make it digestible and actionable.