What it does
In plain words.
A mentor sets up a group, students upload their drafts, and each one comes back scored with what to fix before filing.
- 01
No account for students
Mentors create groups and share a public link; students submit through it without signing up.
- 02
A score out of 100
Eleven weighted dimensions, from inventive step and novelty to claims and commercial potential, plus strengths, weaknesses and a refined inventive concept.
- 03
Citations and copying, checked
Every patent a draft cites is looked up in live patent search, and the text goes through a plagiarism scan.
- 04
Built for mentors
A dashboard with analytics, every submission, a live view of the pipeline, retries, Excel export and PDF reports.
How it works
Under the hood.
Context first: the patents a draft cites are checked against live databases before the model grades a word of it.
A deterministic stage runs before the generative one, so the model reasons from verified patents, not from memory.
- 01
Extract
Patent IDs are pulled from the draft with regex and text patterns.
- 02
Verify
Each ID is looked up in live patent search (Serper), giving a ground-truth list of valid and invalid citations.
- 03
Evaluate
The PDF and that list go to the model (Gemma by default), which returns structured JSON, sanitised before use.
- 04
Check
Copyleaks scans for plagiarism, and the evaluator looks for its own canary phrase.
- 05
Persist
Reports land in Cloudflare D1, files in R2; pipeline logs, a dead-letter table and retry counts record every stage.
Decisions that mattered
Edge router, heavy service
A Cloudflare Worker handles routing, auth and storage; an Express service on Render does the slow work: PDFs and model calls.
Deterministic before generative
The citation check runs first, so the model is grounded in which patents actually exist before it judges anything.
Integrity as a formula
+3 for each valid patent cited, up to +15, and −5 for each invalid one, applied by the model alongside a verdict from Excellent to Fraudulent.
A canary in the report
Each PDF report hides an invisible line. Paste it into an AI to rewrite the draft and the phrase comes back: the resubmission scores 0 and is rejected in code.
Exhibit A
The rubric.
Every draft is scored on the same 100 points.
Still curious?