Live · Qwen3.5-2B + LoRA · trained on CUAD

Upload a contract, see the risky clauses highlighted.

A 2B-parameter model fine-tuned to find 41 kinds of clauses lawyers look for, from uncapped liability to change of control. It lifts held-out span F1 from 0.02 to 0.49 over the same model untuned.

Not legal advice. Your file is read in your browser; only its text is sent for scanning, and nothing is stored. How your contract is handled

01— Scan

Drop a contract here, or choose a file PDF (text, not scanned), DOCX or TXT · up to about 60 pages

Or paste the text

Or try a demo contract

02— Ask this contract

Ask a question, get an answer that points at the text.

Scan a contract or open a demo first.

03— How it works

A small model, measured before and after fine-tuning.

Held-out results · 97 CUAD contracts

Metric	Zero-shot	+ LoRA r8
Span F1 (IoU ≥ 0.5)	0.023	0.492
Clause presence F1	0.387	0.833
Unparseable outputs	58.6%	4.1%
Output tokens	709k	175k

Strong: governing law 0.86, agreement date 0.76, parties 0.70. Weak: non-compete 0.26, termination for convenience 0.21. Cards for weak types say “low accuracy”.

Pipeline

In your browser: pdf.js or mammoth extracts the text and splits it into overlapping sections of about 1,500 tokens, with the same chunker the model was trained on.
Cloudflare Worker: builds the fixed prompt and checks the daily budget.
Modal GPU (L4): vLLM runs Qwen3.5-2B with the LoRA merged in. It returns each clause's type and its first and last words.
Back in the Worker: those anchors are matched to exact character offsets, so highlights land on the real text.
Ask embeds sections with bge-m3 and answers with Qwen3-30B-A3B on Workers AI, citing only sections it was given.

How your contract is handled

Your file: Read in your browser. The file itself is never uploaded.
Its text: Sent in sections to a Cloudflare Worker, then to a GPU on Modal (may run in the US) and, for Ask, to Workers AI. Processed in memory and not stored. Modal doesn't store the requests this app sends, and Cloudflare doesn't train on them.
Logs: One line per request: timings, counts, lengths and a keyed fingerprint. Never the text or your questions.
Your IP: Turned into a pseudonymous id, salted afresh each day, to enforce a fair daily quota. Deleted after the day.
Tip: Remove personal details before scanning a real contract.

Cost and limits

GPU spend is capped at $0.23 a day (under $10 a month). When it runs out, live scanning pauses until midnight UTC and the demos keep working. Scans are fast once the model is warm. The first scan after a quiet spell usually waits about 30 seconds for the GPU to start, and occasionally a few minutes.

The model has only been measured on CUAD: US commercial contracts filed with the SEC. Leases, employment contracts and non-US law are untested. Highlights can be wrong or missing.