Made with Fable 5.1Anyone in a meeting who wants to see the conversation while they are in it
Hyperion
A shared web page that listens to a meeting and surfaces what is being missed: misunderstandings, unstated assumptions, drift, and the question nobody looked up.
Live. Click once and you are in a room. Photo: Jordan Encarnacao on Unsplash.
Hyperion's feed on a fictional meeting: transcript on the left, what the model surfaced on the right.
People can only process one stream, what is being said. Hyperion runs the second stream, the meta-conversation, and shows it on a canvas every participant can glance at without leaving the discussion. There are no fixed card types; the model decides what deserves attention.
Four subsystems, four cadences
Fact detection runs on every final transcript fragment with a five-second throttle. Master analysis runs every sixty seconds over the full transcript. Problem framing runs every five minutes. A web-searched deep answer runs on demand. Each picks its own model, context size, and latency budget.
Zero framework cognition
No regex, keyword lists, or rule chains stand in for understanding. If a decision needs meaning, a model makes it. Call-type routing was deleted in favor of one master prompt with a menu of feed item types.
Homemade LLM observability
Every model call writes one row to Postgres with tokens, latency, and cost. A text-to-SQL ask box lets you interrogate those rows in plain language over a read-only role.
A prompt lab for the prompts that ship
Side-by-side runs across models per prompt variant, a merge cell that synthesizes them, timeline analysis in five-minute buckets, and an evaluation harness fed by real transcripts.
Photo: Jonathan Velasquez on Unsplash
Specifications
- Stack
- Next.js 16, React 19, Tailwind v4, Vercel
- Transcription
- Deepgram Nova 3 streaming, word timestamps, diarization
- Models
- Claude Sonnet 5 with web search, Gemini 3.1 Pro, plus Crusoe and NVIDIA providers behind one interface
- Real-time
- Ably pub-sub with token auth per room
- Observability
- Neon Postgres call log, Datadog RUM and LLM Obs
- Languages
- English and French, UI and model output
- Rooms
- Random ID, shareable URL, no accounts, no persistence
Principles
- 1Minimal and glanceable. The feed never distracts unless the value clearly warrants it.
- 2Multiple subsystems at different frequencies, each optimized for its job.
- 3Maximize managed services; run as little infrastructure as possible.
- 4Docs are treated as code. A stale claim fails review like a stale type.