Skip to content
Back to projects

Case study 04 / 18 · Websites & apps

Sevro AI

My project, with contributors: Includes switchable transcription, which I built

Status: In development

A conversation-intelligence platform in development, with an initial workflow for tutoring discovery calls. Its local workspace combines live transcription, call-bound recording, checklist guidance, and post-call review.

Case study

Sevro AI is my project, built with contributors. It is a conversation-intelligence platform whose first use case is a single-user workflow for tutoring discovery calls. The codebase separates the local call workspace, the public marketing site and the local provider and gateway services.

The problem

Help someone lead an important conversation while keeping their objectives in view, and leave them with a structured record afterwards.

What was built

  • A local call workspace with live transcription, call-bound recording and recovery, checklist guidance and post-call review.
  • Switchable transcription, which I built.
  • A separate marketing site with adaptive visuals and an optional local FAQ assistant.

Key decisions

  • Start with one focused use case, tutoring calls, before presenting other sectors as validated.
  • Keep the marketing FAQ assistant separate from live coaching in the call workspace.
  • Offer replay and static, reduced-motion alternatives, and keep providers configurable.
  • Keep local implementation and synthetic testing distinct from hosted-product readiness.

Outcomes

  • Switchable transcription merged on 11 October 2026 (18 commits).

Where it stands

In development and running locally. Hosted sign-in, provider and device qualification, coaching evaluation and release gates are still open, and there is no public signup yet.

How it was built

Frontend

  • React
  • TypeScript
  • Vite
  • React Router
  • Tailwind CSS
  • Three.js — Adaptive visuals on the marketing site

Backend

  • Fastify — Local gateway and provider services
  • WebSocket — Live call data between the workspace and the gateway
  • Python — Native speech helper for local transcription

AI

  • Whisper (faster-whisper, whisper.cpp) — Local speech-to-text engines in the switchable transcription I built
  • sherpa-onnx — Another local engine behind the same switch
  • Transformers.js — Whisper running in the browser on WebGPU or WebAssembly
  • Deepgram and AssemblyAI — Cloud transcription engines behind the same switch
  • Ollama — Optional local FAQ assistant on the marketing site (Qwen3 4B), answering from a Markdown knowledge pack
  • ElevenLabs — Spoken replies and push-to-talk transcription for the FAQ assistant

Infrastructure

  • Docker
  • nginx

Testing

  • Vitest
  • Playwright — Browser tests with axe accessibility checks