Capabilities
One workspace for search, claims, review, and drafting
Patendra keeps the technical disclosure, retrieved sources, claim versions, review findings, and first draft in one matter. Each important conclusion stays connected to the source or rule that produced it, and unresolved issues remain visible for professional review.
Prior art
Prior-art search with source visibility
A search conclusion is only as useful as its source coverage. Patendra can query four configured backends, combine the ranked results, and show when a source failed or returned too little evidence.
USPTO Open Data
Full-text search over U.S. patents and published applications through the USPTO Open Data Portal, the same corpus an examiner starts from.
Espacenet (EPO)
Worldwide coverage through the European Patent Office's OPS service, so art published only outside the U.S. still surfaces before it can surprise you at examination.
Google Patents Public Data
BigQuery-backed search across the Google Patents public corpus for broad recall on older and non-English families.
Local vector store
A ChromaDB semantic index built on your machine finds conceptually similar art that keyword queries miss, and it keeps working offline.
Results from available sources are merged with Reciprocal Rank Fusion and can be re-ranked with a cross-encoder. Only retrieved references with verified identifiers are eligible for use in a rejection. Generated source identifiers are excluded. Read how Patendra searches prior art.
Claim optimization
Claim refinement with structural checks
Claim scope is a trade: broad claims are valuable and fragile, narrow claims are safe and worthless. Patendra treats that trade as a search problem and explores it systematically.
-
Tier 0: structural validation, free
Every candidate claim set first passes deterministic structural checks: antecedent basis, claim dependencies, element consistency, statutory class. No tokens are spent on a claim that fails grammar an examiner would flag in a first Office action.
-
Tier 1: quick anticipation check, cheap
Surviving candidates get a fast novelty screen against the retrieved art. Claims that read directly on a single reference are pruned before the expensive stage ever sees them.
-
Tier 2: adversarial search and full examination
The strongest candidates face a fresh, adversarial prior-art search aimed specifically at killing them, then a full simulated examination. Best-first iteration then refines scope: broaden what survives, narrow what gets rejected, and keep the evidence trail for every move.
Examination
Examination issue review by statute
Pressure-test a claim set against common §102, §103, and §112 issues before a filing decision. The result supports review and does not predict an examiner outcome.
Anticipation
Element-by-element mapping of each independent claim against the closest retrieved references. A §102 rejection is only asserted when a single reference covers every limitation, and the reference is cited by ID.
Obviousness
Combination analysis across references with an articulated reason to combine, so the review records more than a bare assertion that the claim appears obvious.
Enablement & definiteness
§112(a) written-description and enablement checks against the specification, and §112(b) definiteness checks on the claims themselves. These are the rejections that don't need prior art to hurt you.
The dashboard reports findings per statute, never a single "ALLOWABLE" badge, and shows an explicit insufficient-search warning when prior-art coverage was too thin to support a novelty conclusion. "Zero anticipating references found" is only good news if the search actually looked. Details: deterministic scoring & honest reporting.
Scoring
A consistent score with its inputs
The LLM never gets to make up the number. Patentability is computed from measurable signals, so the same evidence always produces the same score.
| Signal | What it measures | Where it comes from |
|---|---|---|
| Novelty | Anticipating references found, and element-coverage of the closest art | Retrieved, ID-verified references |
| Non-obviousness | Viability of §103 combinations against the claim set | Simulated examination memo |
| Enablement | Specification support and §112 validation findings | Structural checks + examination |
| Breadth | Claim scope actually retained after refinement | Claim structure analysis |
Confidence is reported as Strong / Moderate / Weak with the raw value on hover, and a Bayesian evidence chain records why the system believes what it believes. When you record your own assessment of a finished run, the weights recalibrate. See the learning loop below.
Drafting
Editable first-draft assembly
Reviewed matter context carries into editable documents for counsel instead of ending as disconnected fragments.
Full written description
Background written as a prosecution argument, a detailed description drafted against a §112(a) enablement checklist, embodiments, and definitions, all grounded in the claims that actually survived examination.
Optimized claim set
Independent and dependent claims as refined by the tree search, with the per-claim scores and the examination memo that justified their scope.
Patent figures
System diagrams, flowcharts, and component figures generated from the specification, numbered and referenced in the description. No external drawing tool required.
PDF application package
A formatted PDF of the complete application, plus every intermediate artifact (hypotheses, search logs, examination memos, score tables) for the file wrapper.
Learning loop
Local calibration from completed reviews
-
Human assessment feeds calibration
Rate a finished run Strong / Moderate / Weak / Rejected in the dashboard. Patendra logs predicted score against your verdict and recalibrates scoring weights and confidence over time.
-
Search strategies are scored too
The search optimizer records outcomes per strategy: which query formulations found killer art, which wasted budget. It prefers what has worked in your domains.
-
Experiential memory across runs
An on-disk memory of prior runs, hypotheses, and outcomes means the tenth run in a domain starts smarter than the first. All of it stays local, in your data directory.
Platform
A smaller, explicit trust boundary
Runs on your machine
Native apps for macOS and Windows. Your invention disclosures, drafts, and API keys stay in your local data directory. Nothing is uploaded to us. Security model
Live run streaming
Real-time progress over Server-Sent Events: phase stepper, live logs, activity feed. No fake progress bars: the only determinate bar is the refinement loop's real iteration count.
Claude or Gemini
Bring your own API key for Anthropic Claude or Google Gemini and switch per run. The model picker disables providers that have no key configured instead of letting a run fail.
See how Patendra handles a real matter
Bring a representative disclosure to a focused walkthrough. We will show the search record, claim review, unresolved items, and editable draft without hiding the limits.