Chatbot

Problem

The chat platform’s native bots were hard to change and almost impossible to test, and an earlier proof of concept had grown into one long chain of if/elif branches that nobody wanted to touch.

Solution

A TypeScript service that receives chat webhooks and runs each conversation through a declarative flow graph. The engine is a pure function: it takes the graph, the current state and the customer’s input and returns the new state plus a list of actions, with no I/O. A separate executor sends messages, opens categorized helpdesk tickets and closes conversations. One process serves several flows on the same engine.

Engineering highlights

• Nearly 11,000 conversations handled in its first 16 days in production.
• Conversation flows as data: menus and quick replies defined as graph nodes and validated at startup, so a broken flow never goes live.
• Conversation state lives in PostgreSQL, one row per conversation locked with SELECT ... FOR UPDATE, so messages from the same customer are handled in order.
• Webhook intake verifies a signature or token, validates the payload with Zod, answers 200 immediately and drops duplicate events by request id.
• Outgoing calls retry on 429, 5xx and network errors, respect Retry-After and use a timeout per attempt.
• Golden-file tests compare every bot message with the specification documents in both directions. The suite has close to a thousand test cases.
• Analytics record the path each conversation took without storing anything the customer typed.