Personal build · AI
Tafsiri AI — machine translation for Kenya’s languages
Swahili, Somali and English translation on a fine-tunable neural model, with an active-learning loop that improves quality over time.
- For
- Product build
- Type
- Personal product build
- Services
- Build, Automate
- Stack
- React, TypeScript, FastAPI…
The problem
Most of Kenya’s ~68 languages are poorly served by mainstream translation tools, and general-purpose AI translators rarely say where they’re unreliable.
The solution
- Translation between Swahili, Somali and English on a multilingual M2M100 model, with other Kenyan languages listed honestly as a roadmap.
- Domain glossaries (medical, legal, technical) that guarantee chosen terms appear in translations.
- Every translation gets a confidence score; low-confidence results go to a review queue where corrections feed future model improvement.
- A LoRA fine-tuning pipeline for adding new languages from community-contributed parallel text.
- Accounts, history and analytics, a community forum, blog, GDPR data export and erasure, and optional billing.
My role
Designed and built end to end — frontend, API, translation microservice and training pipeline.
Technology
- React
- TypeScript
- FastAPI
- PostgreSQL
- Redis
- PyTorch
- Hugging Face Transformers
- Docker
Outcome
Live as a public demo. A personal product, so there are no commercial figures.
Facing a similar problem in your business?
Start with a conversation about the problem. We’ll work out the right solution together.