professional work

Conversational AI Onboarding

A chat-based onboarding flow that populates a user's AI agent knowledge base through a guided conversation.

Full-Stack Developer · Undisclosed · Oct 2025 — Present

context

New users landed on a blank knowledge base with no guidance on what to put there. The goal was to replace that empty form with a short conversation that asked the right questions and filled in the knowledge base automatically.

the challenge

Getting multiple things to work together: an interview state machine, BAML typed functions, SSE streaming, website scraping, and knowledge base persistence. LLM output is unpredictable and SSE events can arrive split across chunk boundaries.

approach

• 1.I built the interview state machine separately from the UI. It tracks interview stage, captured fields, and completion status. Keeping it decoupled made it easier to reason about and test independently.
• 2.I used AI classification for detecting user intent (confirming, correcting, or continuing) instead of regex. Regex would need constant updating. AI classification handles natural variation without a keyword list.
• 3.I wrote an SSE streaming parser that handles events split across chunk boundaries. It buffers partial events and only emits once the full event is assembled.

technical highlights

• —BAML typed functions for each interview stage: question generation, answer merging, intent classification, and confirmation. Typed functions enforce structured output at the schema level so the LLM can't return unexpected shapes.
• —After the interview finishes, the system scrapes the user's website with a 45s timeout, saves it to the knowledge base, and links it to the created AI agent.
• —The chunk-boundary SSE parser was necessary because the network sometimes splits a single SSE event across multiple fetch chunks, which drops tokens if you process each chunk independently.

result

Users end up with a pre-populated knowledge base after a short conversation instead of a blank form.

technologies