Let's talk
← all work

work/language-ai-platform.tsx — case study

AI · Awarri

Language AI platform

From raw language data to a voice assistant that speaks local languages.

role
Senior Frontend Engineer · frontend team lead
period
2022 – 2023
stack
ReactTypeScriptLiveKitAccess control
X-ray · under the hood
  • ›text and audio data in
  • ›annotators translate text or correct machine transcripts
  • ›QA reviewers: validate · confirm · edit · send back
  • ›roles: annotator · QA reviewer · super admin
  • ›up to 100 annotators and 15 QA reviewers

← drag to peek under the hood →

Illustration, not a screenshot. Drag the handle to see what's under the hood.

the challenge

AI that works in local languages needs clean, carefully checked data. Machine learning engineers needed a dependable way to turn raw text and audio into trustworthy training data across 4 to 5 local languages plus English, and then a way to put the models trained on that data into people's hands.

how it fits together

  1. raw text & audio
  2. annotation
  3. QA review
  4. clean training data
  5. speech & text · voice cloning · voice assistant

what it does

  • Annotation workspace

    Text or audio goes in. Annotators either translate the text into their local language, or correct a transcript that an automatic system got wrong.

  • Quality review

    QA reviewers check every submission. They can validate, confirm or edit it, or send it back to the annotator to rework.

  • Roles and access control

    Three roles, annotator, QA reviewer and super admin, with access control so each person only gets the tools their role needs.

  • Speech and text

    Text goes in and clearly spoken audio comes back, in your chosen language and dialect. Audio goes in and clearly written text comes back, with all its diacritics.

  • Voice cloning

    Record your voice, save it, and use it in other services such as speech translation, text to speech and speech to text, in a local dialect.

  • Real-time voice assistant

    An assistant you can ask questions in local languages. It answers in real time, built on LiveKit, and brings the other products together.

what I did

  • Led the frontend team of 5 and did most of the hands-on work across all three products.
  • Oversaw design, development and deployment of the user-facing features.
  • Built the annotation and QA workflows for text and audio data, including access control for annotators, QA reviewers and super admins.
  • Built the frontends for the speech and text product, voice cloning and the real-time voice assistant.
  • Introduced testing practices and code reviews, reducing bugs and speeding up development cycles.

impact

  • Gave machine learning engineers clean, checked data to train models in local languages.
  • Supported up to 100 annotators and 15 QA reviewers working in the tools.
  • Covered 4 to 5 local languages plus English.
  • The data fed directly into the finished products: speech and text, voice cloning and the voice assistant.
  • Fewer bugs and faster development cycles through testing and code review.

see also

working on something similar?

Let's build it together.

Start a project →