Real Projects with
ElevenLabs

• Twilio number, Deepgram STT, Claude, ElevenLabs streaming.
• Books appointments in Google Calendar, logs calls in Supabase.
• Business model: $499/mo per location.

Uploads a CSV of leads, dials each with Twilio, qualifies with Claude, books meetings on positive intent, sends recap SMS after hangup.

SYSTEM: You are Ava, calling on behalf of Acme Cloud.
Goal: book a 15-minute demo with any decision maker.
Never oversell. If uninterested, thank and end within 30 seconds.

Candidate joins a link, ElevenLabs asks 6 role-specific questions, Whisper transcribes, Claude scores against a rubric and produces a hiring report.

const rubric = ["problem solving", "communication", "ownership", "domain depth"];
const report = await claude.messages.create({
  model: "claude-sonnet-4-5",
  system: "Score each dimension 1-5 with one sentence of evidence.",
  messages: [{ role: "user", content: `Transcript:\n${transcript}\nRubric: ${rubric.join(", ")}` }],
});

Student speaks a question, tutor answers with voice, quizzes back, and adapts difficulty. Persist a per-student skill graph in Supabase.

1. User uploads a topic + 3 bullet points.
2. Claude writes a 2-host script (Alex + Sam).
3. ElevenLabs synthesizes each line; stitch WAVs with ffmpeg.
4. Output: a 6-minute MP3 with intro music.

ffmpeg -f concat -safe 0 -i lines.txt -c copy episode.mp3

Chunk manuscript into paragraphs (max 4000 chars), stream each through ElevenLabs with previous_text/next_text for prosody continuity, stitch into chapter MP3s.

User sends a voice note to your Twilio WhatsApp number. Transcribe with Whisper, answer with Claude, synthesize with ElevenLabs, reply with the MP3.

app.post("/wa/incoming", async (req, res) => {
  const audio    = await downloadTwilioMedia(req.body.MediaUrl0);
  const text     = await transcribe(audio);
  const reply    = await think(text);
  const mp3      = await speak(reply);
  const publicUrl = await uploadToSupabase(mp3);
  await twilio.messages.create({
    from: "whatsapp:+14155551212", to: req.body.From, mediaUrl: [publicUrl],
  });
  res.sendStatus(204);
});

Same skeleton as receptionist, with a RAG tool over the docs. On low confidence, opens a ticket and voices: "I've raised a ticket — a teammate will follow up."

• voice.ts — one function per pipeline (transcribe, think, speak, log).
• supabase/calls — the shape every project logs to.
• lib/ffmpeg.ts — stitch, denoise, normalize.
• middleware/verifyTwilio.ts — signature check.