How siddanth-6365/trackit-ai works
Project type and purpose
• This is a full-stack web application with a complementary Telegram bot.
• Purpose: turn receipt images into structured expense data (OCR → LLM parsing → Supabase), allow natural-language queries over spending, and expose both a browser UI (Next.js) and a Telegram interface.
Main components
a. Frontend (Next.js + Tailwind + shadcn/ui)
b. Backend API (FastAPI + asyncio in Python)
c. Telegram Bot (python-telegram-bot CLI)
d. External AI & OCR services
• Tesseract OCR engine
• meta-llama/llama-4 models API for receipt parsing
• sqlcoder-7b-2 via Cloudflare AI for NL→SQL
e. Database (Supabase PostgreSQL)
Component relationships & data flows
• Browser → Frontend UI → REST calls → Backend API
• Backend routers (/receipts, /query, /users) delegate to service layer
• Service layer modules:
• ocr_service.py calls Tesseract
• llm_service.py calls llama-4 API
• query_service.py calls sqlcoder API then executes on Postgres via supabase_client.py
• receipt_service.py orchestrates OCR → LLM → DB
• user_service.py handles auth & user data
• Backend → Supabase PostgreSQL (data storage)
• Telegram Bot → calls Backend endpoints (/receipts/telegram_upload, /query/ask) → same backend flow
Architectural patterns & design principles
• Layered architecture: Presentation (frontend/telegram), API gateway (routers), Domain logic (services), Data access (Supabase client)
• Async IO for high concurrency in Python
• Separation of concerns: routers vs. services vs. data client
• Stateless backend (all state in DB)