Colab91 — AP Classifier

Built at Sthaan.ai as an AI Engineer. Colab91 is an Electron desktop application that takes an imported CSV of accounts-payable transactions and classifies every row automatically — supplier, spend category, and GL code — using a multi-agent LLM pipeline.

The four agents

The pipeline is four cooperating agents, orchestrated over OpenRouter:

• Column Mapping Agent — maps raw, inconsistent CSV columns onto a canonical schema, so the rest of the pipeline sees a stable shape regardless of how the source file was exported.
• Research Agent — performs Exa-powered web research to enrich each supplier with real-world context before any classification happens.
• Classification Agent — assigns supplier, spend category, and GL code using a taxonomy plus the context gathered upstream.
• QC Agent — validates the classifications and flags low-confidence rows for human review instead of silently guessing.

Why multi-agent

Splitting the work means each agent has one job and one prompt to get right. Mapping failures don't contaminate classification, research is reusable across rows for the same supplier, and the QC pass turns "the model was confidently wrong" into "the model asked a human" — which is the difference between a tool finance teams trust and one they don't.

Stack: Electron · OpenRouter · multi-agent LLM pipeline · Exa.