Shadin A FinChat
Solo product, live
| Measure | Before | After |
|---|---|---|
| Runs that succeeded | 29% | 100% |
| Errors and budget overruns | 183 | 0 |
| Plato’s mission failure rate | 73% | 0% |
| Tokens on a benchmark run (indexed) | 100 | 63 |
Before: the 7 days up to the fixes. After: the next 4.5 days, 91 runs. Plato’s rate and the token benchmark were measured separately.
- What ran
- A seven-agent AI platform that runs scheduled missions every day and sends the results to Telegram. Node.js, PostgreSQL, routing across several AI providers, and bring-your-own-key so AI cost scales per user.
- What worked
- I shipped the fixes in a day. Runs went from 29% successful the week before to 100% over the next 4.5 days, 91 runs in all, and errors and budget overruns fell from 183 to none. Plato’s mission failure rate fell from 73% to 0% after I added a paid provider. On a benchmark run, tokens fell 37% once I found 86% of them went to re-reading the prompt.