# All Vibes

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Pramodh Nanjayya](https://cerebralvalley.ai/u/all_vibes)
- **GitHub:** https://github.com/pramodh-bn/dialogiq-flight-deck
- **Demo video:** https://youtu.be/twqMTnczoek
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2730

DialogIQ: a flight simulator for hard conversations

INSPIRATION
Most managers were promoted for doing the job, not for hard conversations. Pilots train in simulators because crashing is expensive. But the conversation that decides whether a good person stays or leaves gets one take. DialogIQ lets you practise it first.

Our original vision was a live voice conversation: Bob speaks, and the coach guides your tone and delivery as well as your words. We built the text-based core first so the character, coach and scoring were solid.

WHAT IT DOES
DialogIQ is a rehearsal app for difficult workplace conversations. Pick a scenario, such as Bob, a guarded engineer who missed his Q3 deadlines, and talk to him in your own words.
- A believable character: Bob has hidden needs he won't just hand over. Pressure keeps him guarded; acknowledgment and real questions open him up.
- One coaching hint per turn: an approach, an example line and a reason.
- Rewind and branch: the Flight Tree saves every turn, so you can go back to the moment before things went wrong and try another line.
- Visible progress: Bob's stance (guarded, warming, open) and how many of his 7 hidden needs you've uncovered.
It rehearses the conversation. It never records or scores the real one.

HOW WE BUILT IT
- Bob runs on Meta's Muse Spark Open through Baseten.
- The coach also runs on Muse Spark Open, through Fireworks. It never sees Bob's hidden needs, and Bob never sees the hints.
- A deterministic judge scores needs uncovered; an opt-in Muse Spark evaluator classifies each exchange with cited evidence.
- FastAPI + SQLite store the turn tree; a TypeScript frontend runs as a Cloudflare Worker. Keys stay server-side.
- Built by indie developer Pramodh, leading a team of AI coding agents (Claude, Codex and Meta's Muse Code) coordinated on a Megillah board in planned sprints.

WHAT WE MEASURED (small, agent-graded tests)
- After the coach redesign, usable hints rose from 23 to 52 of 78.
- When Bob asked "Can we stop?", every hint said to respect it: 6 of 6.

CHALLENGES
- Making the coach useful without giving it the answers.
- Credit: our $150 Muse Code token credit ran out overnight with no warning, and we paid out of pocket to finish.
- Latency: the median first hint takes about 14 seconds on Fireworks, which misses our goal.
- Honesty: the stance is a simulated state, not a measure of real trust, and we document known misses.

WHAT WE LEARNED
A good coach shouldn't know the answers. It should know you. Keeping the persona, coach and judge separate made each easier to test.

WHAT'S NEXT
Live voice conversations with tone and delivery coaching, better coaching after the learner's own mistakes (5 of 12 today), faster hints and more characters.

Built with: Meta Muse Spark Open, Baseten, Fireworks AI, FastAPI, SQLite, TypeScript, Cloudflare Workers

## More from Meta Global AI Developer Hackathon

- [Unknown Ally](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2727)
- [Lusong](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2728)
- [Azmuth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2729)
- [David Koch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2731)
- [wetoyo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2732)
- [Apex Business OS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2733)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2730. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt