LLM Strategy

Hybrid Approach

• Simple tasks: Regex classification → deterministic handler (0ms, zero cost)
• Complex tasks: LLM agent with function-calling (2-3s per turn, ~$0.01/task)

Primary: Gemini 2.5 Flash (Vertex AI)

• Fast (~2s/turn), native function-calling
• Available via existing GCP account

Fallback: Claude Sonnet 4.6

• Better reasoning for complex tasks
• PDF understanding for invoice OCR

Agent Loop

• Max 15 turns per task
• Auto-fix 422 errors between turns
• Pre-fetch company context during first LLM call
• 280s timeout