Labrys Discovery Bot:
AI Discovery and Proposal Generation

Discovery is the highest-leverage part of a services sale and the most repetitive — standard calls burn 30 to 60 minutes pulling out the same baseline information every time. Scripted chatbots feel robotic: they restate understanding, then ask the next item off a list, and the prospect tunes out. A real discovery bot needs to press on vague answers without becoming an interrogation, sound like a commercially confident senior rather than a checklist, and still produce output structured enough to generate a real proposal from — not notes someone has to rewrite.

Answer classification and conversational response cannot share an LLM call — mix them and the classifier gets soft or the conversation gets robotic, so they have to run as separate concerns. Context windows are finite: a 20-turn discovery blows past token limits if you send the full history each call, forcing summarised and windowed context. Serverless execution limits mean proposal generation — 30 to 60 seconds of LLM work — cannot block the HTTP response, so it must run after the page returns. Deep conversations that fail to resolve need a soft landing via depth capping, partial briefs, and a Calendly handoff when the bot is the wrong tool for the job.