| Backtester | Every number: the engine runs in your browser, with Dhan delivery charges, next-day-open fills and a run hash. Verify runs 246 real backtests. | Prices are sample prices unless the deploy fetched NSE daily candles (it tries at build time). The rule reader is code, not a model, unless DVAR_API_KEY is set. |
| White-label service | Nothing: it is the proposal. | API shape, service levels, model names, timings. |
| Dhan Assist | Routing, the five runtime checks, the cost arithmetic; portfolio P&L, sector shares and option payoffs computed from the sample account. | The sample account (holdings, positions, orders), help text, quotes, news lines and answers, and the 60-day price paths in the charts (seeded, illustrative). The experience is shaped after Groww's GR1 (public reports from its March 2026 beta); its exact UI is not public, so the card layout is ours. No model is called unless DVAR_API_KEY is set. |
| Evals: scorecard | All seven checks and the release gate are code and run in your browser. | The 30 chats and the three answers to each, written to show typical behaviour. Costs are list-price assumptions at 4 characters a token. |
| Evals: Eval Console | Scores, intervals, topic minimums, the four gates, judge agreement and kappa, and costs from token counts, computed as the eval engine computes them. | The 180 chats, the replies, which reply fails which check (drawn per model from set failure rates, seeded), latencies and the judge's verdicts. The fine-tuned model was not trained. The frontier price is a list-price assumption; the Gemma prices are Dvar Inference's. |
| Post-training | The rule-writer score of the built-in reader on 40 sentences, field by field. | No model was trained; there is no tuned score yet. Model names are placeholders. |
| Document scanning | Schemas, format, masking, date and sum checks, name and date-of-birth matching across documents, routing, review queue, scoring against the answer key. | The five specimen documents (real layouts, fictitious person and bank, SPECIMEN watermark; emblem, logos and QR data left out), what each model read and its confidence. Per-page prices are assumptions. |