Conversational AI assistant as a Telegram bot for a Ukrainian consultancy helping small businesses apply for government and EU grant programs.
Problem: 1 application = 2–4 hours of consultant time. 50+ clients/month.
Architecture:
- RAG + LLM classifier (pgvector HNSW + OpenRouter) routes to correct program
- LLM function-calling extracts facts from 10–20 client messages
- Donor-specific rule validation (funding limits, min employees, eligible categories)
- Excel template filling — input cells only, formula cells untouched
- Operator escalation with full context when AI confidence drops below threshold
Stack: Go · PostgreSQL · pgvector · Ollama bge-m3 · OpenRouter · Telegram Bot API · excelize · Docker
Problem: 1 application = 2–4 hours of consultant time. 50+ clients/month.
Architecture:
- RAG + LLM classifier (pgvector HNSW + OpenRouter) routes to correct program
- LLM function-calling extracts facts from 10–20 client messages
- Donor-specific rule validation (funding limits, min employees, eligible categories)
- Excel template filling — input cells only, formula cells untouched
- Operator escalation with full context when AI confidence drops below threshold
Stack: Go · PostgreSQL · pgvector · Ollama bge-m3 · OpenRouter · Telegram Bot API · excelize · Docker