Soffia — AI Academic Assistant
Presentations, papers and tests — generated in seconds, for 20K+ users.
An AI assistant that generates presentations, course papers, reports and tests in seconds — on the web and inside Telegram.
- Role
- Senior backend engineer
- Impact
- 20K+ users · documents in seconds
- Live
- soffia.uz ↗
- Status
- In production
- Year
- 2025 – 2026
The problem
Turn a one-line prompt into a finished, correctly-formatted academic document in seconds — from either a browser or a Telegram bot.
Turning a one-line prompt into a finished, correctly-formatted academic document is slow, heavy work — and users expect it in seconds, from a browser or from inside Telegram, without ever knowing what's happening underneath. Users forgive a spinner. They don't forgive a wrong document.
What I built
FastAPI generation services wrap LLM and voice (STT/TTS) models, with NLP automation behind them. The slow generation runs as async jobs (Celery + Redis) so the API answers immediately; one engine is fronted by both a web platform and a Telegram bot. Google and Telegram OAuth keep sign-in friction near zero. The bot and the website are the same engine wearing two coats.
Key decisions
Generation never blocks a request — it's a job, with a status. One generation engine, many front doors (web, bot), so the product can meet students where they already are. Make the model's slowness the queue's problem, not the user's. Async isn't an optimisation here — it's the only honest design.
One generation engine behind two front doors; the slow LLM work runs as async jobs so the app always answers instantly.
20K+ users · documents in seconds
Built with
Lessons from the bench
- — Users forgive a spinner. They don't forgive a wrong document.
- — The bot and the website are the same engine wearing two coats.
- — Async isn't an optimisation here — it's the only honest design.