Tanish
Vadel

Highlights

• Times every turn and splits it into speech-to-text, model and speech, live in the UI.
• Turns that take 1.5 s or more go red, on screen and in the logs.
• Starts speaking on the model’s first streamed words, not after a full sentence.

How it works

Why I built it this way

• Groq’s Llama 3.1 8B as the default model

  It’s the fastest option; Claude follows a sales script better but is slower.

• A 50 ms speech buffer

  Cartesia’s default 3 s buffer would make the agent sound slow.

Highlights

• Hears English and Spanish in one sentence, the way learners actually speak.
• Grades answers by meaning: a free exact check first, then a small AI judge returning strict JSON.
• Remembers your weakest words and brings them back.

How it works

Why I built it this way

• Pipecat over LiveKit

  A one-person prototype needs full control of the pipeline and per-stage timing.

• LangGraph only when a tool fires

  It stays out of the audio path, protecting the 1.5 s reply target.

Highlights

• Dials out over a real phone line (SIP) with LiveKit Agents.
• Asks one question, confirm, reschedule or cancel, and logs the answer.
• Hangs up only after its goodbye finishes, so no one is cut off mid-sentence.

How it works

Highlights

• One speech-to-speech model over WebRTC, so tone and pauses survive.
• A fixed phrase check opens a crisis line (Tele-MANAS 14416), so the model is never the only safeguard.
• A latency radar shows how fast each reply starts: green under 0.5 s, amber under 0.9 s.

How it works

Why I built it this way

• Speech-to-speech instead of speech → text → speech

  A text chain flattens tone and pauses, and adds delay.

• A fixed safety check beside the prompt

  Never let a probabilistic model be the only thing between a person and help.