Finch: a support assistant that hands over instead of guessing

Status: in development. No production users and no client deployments yet, so there are no results to report.

The problem

Most support bots fail in the same place. They answer confidently from nothing, and when they finally give up, the customer starts over with an agent who has none of the context.

Seven years on Upwork's support platform taught me that the parts deciding whether a support bot works are rarely the model. They are what it can retrieve, which model a request is worth, how it hands over, and whether anyone can see how well it is doing.

What I'm building

• Retrieval from the team's own history. Hybrid retrieval over help-centre articles and past tickets, on PostgreSQL with pgvector.
• Routing by cost. A LiteLLM gateway sends each request to a frontier or local model depending on what it needs.
• Handover, not a guess. When retrieval can't support an answer, Finch opens a Zendesk ticket with the conversation attached, so the agent starts with the context.
• Measured, not assumed. Langfuse tracks evaluation and observability across every step.

Why it exists

Finch is the Upwork lesson turned into a product: get the wiring right, and measure it, before worrying about the model.