Build an AI harness that turns a strategic prompt into a fully-featured Rasa bot — then let it help Lisa pick where the whole company flies next.
A harness is a reusable, AI-orchestrated system. Feed it a business use case; it ships a production-ready Rasa bot.
One well-crafted prompt describing the use case, the flows, and the goals. No hand-coding the bot.
Agents, an LLM prompting layer, integration patterns, and human checkpoints working together as an architecture.
CALM flows, dialog logic, error handling, clarification loops, and escalation — live over ngrok or a deployed endpoint.
Documented code + README explaining architecture and design decisions, with example skill definitions.
Use Case → Harness → Bot, annotated with agents, rules, decision points, and human checkpoints.
3–5 real checkpoints where a human validates or adjusts harness output — not just checkboxes.
The day runs like a flight schedule. Tap any row for the gate notes.
Director of HR. She needs to decide where Rasa holds its next company offsite — and she wants an agent to do the heavy lifting.
Click a question to load it into the bot demo below ↓
Amadeus, Kiwi Tequila, SerpApi, Booking.com, Google Places — pull live prices and availability.
VisaGuide.World, IATA Travel Centre, government advisories — requirements by nationality.
Open-Meteo, WeatherAPI, TripAdvisor, Google Maps — climate by month and group-friendly things to do.
Tap a stage to inspect it. The middle is where your team's point of view lives.
The offsite brief + goals fed in as a strategic prompt.
Your architecture that plans, generates, and validates.
Flows, actions, escalation — live endpoint.
A taste of the conversation your bot will need to win. Pick what Lisa asks.
An LLM judge reads each exchange and asks one thing: was the answer correct, useful, and well-reasoned? Handling trade-offs and edge cases is where points live.
LISA ASKS →
One number decides the leaderboard. Everything else is a tiebreaker.
After 7:00pm, every bot is run against 50–100 labeled conversations. Each specifies a user message, an expected action, and a desired outcome. An LLM judge scores whether your bot achieved the stated goal.
% of conversations where the bot's output matched the expected outcome. Ranked descending.
% of flows your harness successfully generated.
Everything below is due by cutoff. Check items off as your team ships them.
BUILD SOMETHING LISA WOULD TRUST · GOOD LUCK, TEAMS