Transparency
How our AI works, and what it never does
What the AI does
- Reads photographed documents and pulls out the institution, number (masked), names, dates and amounts, with a confidence per field.
- Argues for and against a claim route (the AI Review) and writes a verdict with reasons and risks.
- Understands your spoken or typed questions.
What it never does
- It never decides the claim route. A fixed, tested rule engine does.
- It never submits anything to a bank or portal. You do.
- It never lowers a caution: the referee can only add risks or ask for human review.
- It never sees your passwords or OTPs.
- Your documents are not used to train any model.
Models
Which models and why
Claude Opus 5 (claude-opus-5) through the official Anthropic SDK for document reading and the debate, because it reads Indian-language documents well and supports strict JSON schemas (structured outputs), which keeps the app safe from unexpected answers. Adaptive thinking is on. System prompts are cached, so repeated calls cost less. Speech in and out use the browser's own engine, which costs nothing and works offline.
Test results
Our evaluation set
We test the parts that must be right. The and the parser are tested by code and run on every change. The AI rows come from a scripted run on the sample documents and need an API key.
| Test | Cases | Result | Note |
|---|---|---|---|
| Rule engine: correct route | 24 scenarios across 5 institutions | 24 of 24 | Every has tests for , no nominee under and over the threshold, joint holder and unknown nominee. |
| AIS parser: rows found | 3 synthetic statements | 13 of 13 rows | Interest, dividend and mutual fund tables. |
| Dividend list parser: real published lists | 3 company lists, 1263 pages, downloaded at check time and never stored | 46,282 rows parsed, search finds 100 percent of a 1128 row sample | Real unpaid-dividend lists that companies must publish under the Companies Act. Counts only; no name is kept. On the 974 page list the company prints its own total on the cover, 49,52,460.80 rupees; the parser sums 49,52,209, so it misses about two rows in 23,709. Reproduce with npm run check:dividends. |
| Claim pack builds | 5 institutions | 5 of 5 | PDF opens, fields filled, three pages. |
| Accessibility (axe, WCAG 2.2 AA) | 11 pages, light and dark | 0 serious or critical | Also zero moderate issues. Reproduce with npm run audit. |
| Lighthouse, landing page, live deployment | 2 profiles, 3 runs each | Desktop 99 to 100, mobile 90 or better | Performance. Accessibility, best practices and SEO are 100 on both. Mobile has ranged from 90 to 96 across runs, because on a throttled connection the score turns on whether the webfont arrives before the largest paint, so the floor is what we claim. Layout shift is 0 to 0.024 against a 0.1 threshold, from the hero stat card moving when the webfont reflows the heading above it. Mobile is 's throttled phone, which is closer to the families this is for, so the lower number is the one that matters. Reproduce with npm run audit. |
| Document reading: institution, number, holder, nominee | 3 sample documents, 12 fields | 12 of 12 | Measured with npm run eval on 2026-09-21. |
| Debate: Challenger finds the planted problem | 3 scenarios with a known problem | 3 of 3 (a single reviewer found 3 of 3) | Name mismatch, amount at the threshold, lapsed policy. On these cases the found no more than a single reviewer did. We publish that because it is the result. The case for the debate is that a family can read the disagreement and that the Referee can only raise caution, not accuracy on cases this clear. Measured on 2026-09-21. |
Last run: 2026-09-21. Run npm test to reproduce the rule engine and parser results. The AI rows need an API key.
Limits
Known limits, in plain words
- Faded or handwritten papers may be read with low confidence. Every field can be corrected by you.
- We have rule files for five institutions. Others get a "needs review" route until their file is added.
- The unpaid-dividend index in this prototype is sample data.
- Hindi and English only, for now.
- The debate improves transparency; it is not a guarantee of correctness. Research is mixed on how much debate improves accuracy, which is why we measure it.
Data
What we keep and for how long
In this prototype, everything you enter stays in your browser. A photo is sent to the AI only when you press the button and is not stored on our servers. One button in the Help tab deletes everything. If you use your own AI key, it stays in your browser too. To report a problem, use the contact link in the footer.