Dhan × Dvar · brief
Five workstreams, one way of working
Dvar builds and runs AI for Dhan and the Raise group under Dhan's name: a verified backtesting service, Dhan Assist, evals on the AI already in production, post-training for Artham, and one harness for small apps like document scanning. Models write words and structure; code does the maths and the checks; a cheaper model ships only through a release gate. About 24 weeks, starting with the backtester and evals because they need the least from Dhan.
| # | Workstream | What Dhan gets | First result | Dvar effort |
|---|---|---|---|---|
| 1 | Backtester outsource | A managed, verified backtesting service for Dhan, Dhan Algos and Stratzy, plus custom models Dhan presents as its own. | Week 6: every daily strategy re-verified | 40 person-weeks to F&O, then about half an engineer |
| 2 | Customer chatbot | Dhan Assist, an investing co-pilot that knows the customer's portfolio, explains stocks and F&O, helps with orders and support, in English, Hindi and Hinglish. | Week 8: internal pilot | 24 person-weeks |
| 3 | Evals on existing use cases | Traces and pass rates on fuzz, ScanX and support traffic, and a gate on every model change. | Week 2: first eval report | 8 person-weeks, then ongoing |
| 4 | Post-training Artham | Task-tuned Artham versions that beat the current one on Raise's own checks. | Week 14: first tuned release | 18 person-weeks plus GPU time |
| 5 | Harnesses for ad-hoc apps | One shared harness for document scanning and similar apps: schemas, checks, review queue, traces. | Week 6: document scanning on the harness | 8 person-weeks |
What the demo shows
- • Backtester outsource: Type a rule, get a result in seconds; verify 10 listed algos with 246 backtests; see the white-label API and service levels.
- • Customer chatbot: Ask about your portfolio, a stock or an option; get cards and charts, follow-up suggestions, orders drafted for confirmation and advice declined.
- • Evals on existing use cases: A 30-chat scorecard in the browser, and the Eval Console: 180 chats, three models, five checks, one release gate.
- • Post-training Artham: A measured rule-writer eval in your browser, the misses turned into training data, the gate and a model card in Raise's name.
- • Harnesses for ad-hoc apps: Five specimen KYC and income documents (PAN, masked Aadhaar, cancelled cheque, bank statement, contract note): field reads, checks, routing, a review queue.
To start
- • A go-ahead on workstreams 1 and 3 as the first commitment, and one product owner on Dhan's side.
- • Read access to historical price data, and the licence terms for using it.
- • A week of fuzz or support traces under NDA.
- • Access to the document scanning app and a few sample documents.
- • A decision on where it runs: Raise's AWS account or Dvar's India-hosted stack.
Concept demo by Dvar. Synthetic data; no Dhan customer data. Effort figures are estimates before seeing Dhan's data, stack or volumes.