Pitch Coach

Practice a pitch out loud and get structured feedback for the type you chose (capital, education, innovation, or technology): a rubric, a score, filler words, and a verdict you can listen to when you want. The interface and the analysis run in Spanish or English (LATAM-first).

Open source: use it, fork it, and keep going.

License: MIT.

Live demo: https://pitch-coach.focampo.com.

Product and implementation docs: docs/README.md.

What a session does

Main loop: choose language, type, and duration → record (MediaRecorder, automatic cutoff) → the server transcribes (ElevenLabs Scribe) → filler words are counted → the model judges the take against the rubric (Nebius by default) → the dashboard shows the score, covered and missing points, and the highlighted transcript.

Same session, after the score:

• Downloadable script with a timestamp per sentence ([mm:ss.cc]) when the pitch was recorded (not when it was typed).
• Listen to the verdict (ElevenLabs; SpeechSynthesis is the fallback).
• Public room — one sourced objection on the first missed point. The room follows the type: investment panel, classroom, innovation committee, or technical buyer. The five rubric ids do not change.
• Figures you said — if the pitch already stated a figure, Tavily looks for a source of the same order. If none is found, the dashboard says so and does not penalize the point.
• Where the time went — when Scribe returned word timestamps, covered points are placed on the audio.
• Second take — 45 seconds on one missed point. If it is covered, this browser remembers the point id and does not store the retake text.
• Ultra analysis — opt-in re-analysis with an extended-reasoning trace.
• Resolve findings — up to 3 follow-up questions on missed points (voice or backup text).
• Cited figures (Tavily) for up to two missed rubric points, with a sentence ready to say out loud.
• Your progress — local history (last 20 practices in this browser), including which points closed or opened versus the previous practice of the same type and language.

The live indicator is a ring that reacts to the voice and sits with the status lines (Listening… / Transcribing… / the closing line), in the session language. The session is anonymous. There is no login.

Screenshots

Pitch setup	Coach (indicator)	Result dashboard

Known limits

• Unit tests, not UI tests — npm test (vitest) covers src/lib/ and API routes (mocked fetch). The microphone loop is checked by hand.
• Anonymous session — history lives in localStorage. There are no accounts and no sync across devices.
• STT — needs ELEVENLABS_API_KEY (Scribe). Without it, or without a microphone, there is a backup text field.
• Transcription is not live — record → stop → transcribe the whole clip.
• Visual coach — a live ring (no glowing orb) that reacts to the audio and settles as the score. Status lines sit with it.
• External providers — Nebius, ElevenLabs, and Tavily have rate limits and need a network. TTS falls back to SpeechSynthesis if ElevenLabs fails.

Stack

• Next.js 16 (App Router + API routes), React 19, Tailwind CSS 4
• Nebius Token Factory — default analysis (MODEL_PROVIDER=nebius)
• Gemini API — manual fallback (MODEL_PROVIDER=gemini)
• MediaRecorder + ElevenLabs Scribe — capture and server-side STT
• ElevenLabs — TTS (verdict and finding questions)
• Tavily — optional dashboard enrichment
• Sentry — errors and tracing (optional; see .env.example)
• Railway — deploy (Dockerfile is production only)

How the models run

Token Factory (NVIDIA, on every practice). The server calls https://api.tokenfactory.nebius.com/v1/chat/completions from src/lib/proveedor-nebius.ts. Nemotron is used in three sizes:

• Take analysis: nvidia/nemotron-3-super-120b-a12b (Super).
• Ultra analysis, only if the user asks: nvidia/Nemotron-3-Ultra-550b-a55b.
• Sparring, entities, the search query, suggestion checks, the room objection, the spoken-figure check, the timeline, and the second take: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B (Nano).

The model does not set the score. The rubric lives in src/lib/rubricas.ts; the server derives the score and counts filler words. Detail: docs/guia-integracion-nebius.md.

Tavily. Search and extract run on the server (src/lib/tavily.ts, POST /api/enriquecer). The transcript is not sent. At most two points per practice receive a figure (MAX_PUNTOS_ENRIQUECIDOS). The same call brings one room objection for the type and checks one figure the pitch already said. If a missed point has no usable figure, there is no suggestion. If a spoken figure has no source of the same order, the dashboard says so and does not penalize the point. Detail: docs/guia-integracion-tavily.md.

The app calls Token Factory at runtime. Gemini remains a manual fallback (MODEL_PROVIDER=gemini) and is not the demo path.

Local setup

Native development on Linux. Docker is not part of the daily loop.

1. Clone the repository.
2. npm install
3. Copy .env.example to .env.local and set at least:
   • MODEL_PROVIDER — nebius (default) or gemini
   • NEBIUS_API_KEY — required with Nebius (the default). Token Factory key.
   • GEMINI_API_KEY — required only if MODEL_PROVIDER=gemini
4. Optional, for the full loop:
   • ELEVENLABS_API_KEY — STT + TTS
   • ELEVENLABS_VOICE_ID_* — Spanish and English (see .env.example)
   • TAVILY_API_KEY — dashboard suggestions
5. npm run dev

A current browser with a microphone. HTTPS outside localhost for the microphone.

Useful commands: npm test, npm run lint, npm run build.

Deploy (Railway)

Build with Dockerfile (Next.js standalone) and railway.toml. The Dockerfile is for this deploy only. It is not used for local development. In Settings → Variables, set the same keys as .env.local (including NEXT_PUBLIC_SENTRY_* and the Dockerfile ARGs if you use Sentry — see docs/sentry.md).

Documentation

Document	What it is for
docs/README.md	Index (tutorial, guides, reference)
docs/alcance.md	Product, business rules, contracts
docs/status.md	What is implemented and where it lives
CONTRIBUTING.md	Contributing and extension points
CLAUDE.md	Rules for agents in this repo

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md.

License

MIT — see LICENSE.