Aurabear AIfrom group-chat chaos to a plan
A text-based AI group planner that turns a chaotic group chat into a booked plan.
The challenge
Aurabear's entire product lives inside the group chat: a dozen people, half-formed ideas, conflicting constraints, and a constant stream of "anyone free Saturday?" Their first agent could hold a conversation, but it guessed wrong too often, cost too much per thread to scale, and slowed to a crawl whenever a chat got busy. To launch, they needed it accurate, cheap, and ready for real volume, all at once.
What we built
We rebuilt the planning agent from the conversation up and put hard evidence behind every change.
- Re-architected the agent as a tool-using automation that reads the whole thread, tracks each person's constraints, and proposes concrete plans: real times, places, and booking links.
- Stood up an eval framework: a golden set of real-world group threads scored on plan-intent accuracy, so every prompt and model change ships on evidence instead of vibes.
- Moved the system onto serverless infrastructure that scales to zero between bursts and fans out under load, collapsing idle cost while holding sub-second replies.
The outcome
Plan-intent accuracy climbed to 94%, cost per planned outing dropped 68%, and the agent now handles 10k+ active group threads a day with median replies under 700ms. Aurabear went from a "maybe it works" demo to a product they could confidently launch.
“Red Crayon took our 'maybe it works' demo and made it something we could actually launch. The eval harness alone changed how we ship.”