BiFL — SaaS platform for AI assistants
My own platform: a catalogue of six model categories, a RAG knowledge base, publishing channels — web widget, Telegram, VK, MAX — and billing. Runs in production.
- AI agents
- Automation
- DeepSeek
- Docker
- ElevenLabs
- FastAPI
- FastEmbed
- Gemini
- GigaChat
- OpenAI
- OpenRouter
- PostgreSQL
- Python
- Qdrant
- RAG
- Redis
- Whisper
- YandexGPT
The brief
Build a commercial platform where customers work with generative and multimodal models from different vendors through one interface — from chat and document search to image, video and audio generation — and pay for actual usage.
BiFL is a SaaS platform for AI assistants of my own: a client area, a model catalogue, a knowledge base and publishing channels — web widget, Telegram, VK, MAX and a mobile app. OpenAI, Gemini, GigaChat, YandexGPT, DeepSeek and OpenRouter models connect platform-side or with the client’s own key (BYOK).
The RAG subsystem extracts text from PDF, DOCX, XLS and CSV, splits it with one of four chunking strategies (fixed, sentence, paragraph, semantic), embeds it with OpenAI, GigaChat or local FastEmbed and searches it in Qdrant. Chunking and embedding run through queues, and every document moves through an explicit lifecycle.
Dialogs are multimodal: voice messages are transcribed on the way in (Whisper, T-Bank VoiceKit) and answers are read aloud (ElevenLabs). Business modules sit on top of the dialog core: catalogue and orders with tool calling from chat, online booking and product recommendations out of the vector index. Billing covers subscriptions and pay-as-you-go.
Outcome
A platform in production: six model categories, a RAG subsystem with vector search, async queues, voice dialogs, BYOK, billing with subscriptions and pay-as-you-go.