About ClauseWise
A careful first read, for people who cannot afford one
ClauseWise helps people understand, compare and act on the legal documents they are asked to sign. It gives information, not legal advice, and it shows its work.
The problem
Priya has been offered a flat. The owner sends a nine-page agreement on WhatsApp and wants it signed tonight. Buried inside: a six-month deposit he may keep “in his sole discretion”, eviction on fifteen days' notice while she is locked in for six months, and rent that can rise 15% with a week's warning.
She is not careless. A lawyer for every agreement is simply out of reach, and nothing tells her which of these clauses are normal. Most rental agreements, job offers, NDAs and loan papers are signed exactly like this.
How it works
- 1
Read
Your file is checked (type, size and real file signature), then its text is extracted.
- 2
Protect
Emails, phone numbers, Aadhaar-like and PAN-like numbers and account numbers are replaced with tokens.
- 3
Analyse
One request asks a language model for the brief, clauses, risks, deadlines and prep pack, in a strict format.
- 4
Check
Every quote is matched word for word against your document. Anything that cannot be found is dropped.
- 5
Score
The balance score is calculated by ClauseWise from the verified clauses, using a fixed formula.
After that, everything else builds on the verified result: grounded questions and answers, comparison with another version or a fair baseline, a negotiation kit, and a prep pack for a lawyer.
Why you can trust what it shows
Personal details masked first
Structured identifiers are removed before any text leaves our server, and restored only in what you see.
Verified evidence
Claims are shown only if their supporting quote exists in your document, with its exact position.
Deterministic scoring
The model labels clauses; code computes the score. The same labels always give the same number.
No storage
Documents live in memory for one request and in your browser tab. There is no database.
Prompt-injection defence
Document text is treated strictly as data. Hidden instructions are ignored, counted and reported to you.
Built to keep working
Free model tiers run out, and a tool that fails when someone needs it is not accessible. ClauseWise uses Google Gemini (gemini-2.5-flash) first, then other Gemini models that each have their own allowance, then Groq (openai/gpt-oss-120b) as an independent last resort.
Whichever model answers, its output passes the same checks: strict format validation, personal-detail restoration and quote verification. If every model is unavailable, the sample documents show a clearly labelled saved analysis; your own documents never do.
Limitations, honestly
- It can misread a clause. Verification proves a quote exists, not that the interpretation is right.
- It does not check the law where you live, or know your circumstances.
- Scanned PDFs without a text layer are not supported yet.
- Names and addresses are sent to the model provider, because they are needed to explain who owes what.
- Very long documents are shortened in the middle; the beginning and end are always read.
Accessibility
ClauseWise aims for WCAG 2.1 AA. It is fully usable by keyboard, works with screen readers, respects reduced-motion and dark-mode settings, fits screens down to 360 pixels, and never uses colour alone to convey meaning. Explanations are available in English, Hindi and Telugu, with a plain-language mode.