Background: NM i AI Competition Learnings
This section documents learnings from NM i AI 2026 (Norway's National AI Championship) that influenced the design of this bot. These are historical notes, not part of the current architecture.
Key Takeaways Applied
| Learning | How It Shaped the Bot |
|---|---|
| LLM agent with function-calling beats hardcoded handlers | Bot uses LLM to decide which API calls to make |
| Pre-fetch common data during LLM call | Bot caches learned patterns in Postgres |
| Auto-fix API errors between turns | Bot retries with corrections on 422s |
| Cheap models are good enough for routine tasks | Gemini Flash primary, not Opus/Sonnet |
| System prompt rules are simpler than a rules engine | Company rules live in the system prompt |
Tripletex Notes (Future Reference)
If Tripletex integration is ever needed:
- Use
/incomingInvoice?sendTo=ledgerinstead of/supplierInvoice - Skip the salary API (always fails), use manual vouchers
- Don't use the bank reconciliation endpoint, register payments individually
- Hardcode VAT type IDs (they are stable across sandboxes)
- Strip known-bad fields (
amountCurrency,voucherDate,isCustomer,isSupplier)
These learnings come from studying 22 teams' implementations during the competition.