AI Project Tracker (Self-Hosted LLM) 🧠

A full-stack task management ecosystem powered by local AI inference.

This platform integrates a Telegram Bot, Web Dashboard, and a FastAPI backend to help users generate project roadmaps and track progress. Unlike standard wrappers, this project runs a local LLM (Qwen 2.5 via llama.cpp) completely offline/self-hosted using Docker.

🏗 Architecture

The system relies on a microservices-like architecture orchestrated via Docker Compose:

graph TD
    User((User))
    TG[Telegram Bot]
    Web[Web Dashboard]
    API[FastAPI Backend]
    DB[(PostgreSQL)]
    Redis[(Redis Cache)]
    LLM[Local LLM Service]

    User --> TG
    User --> Web
    TG --> API
    Web --> API
    API --> DB
    API --> Redis
    API -->|Prompt| LLM
    LLM -->|Stream| API

Loading

🚀 Key Features

Self-Hosted Intelligence: Uses llama.cpp server to run quantized models (GGUF) locally. No OpenAI API keys required. Omnichannel: Manage tasks via Telegram Bot (Aiogram 3) or Web Admin Panel (Next.js). AI Generation: Automatically breaks down a vague idea into 5-7 actionable technical tasks with acceptance criteria. Reactive UI: Real-time progress bars for AI generation on both Web and Telegram.

🛠 Tech Stack

Frontend: Next.js 14, Tailwind CSS (v4), TypeScript.

Backend: FastAPI, SQLAlchemy (Async), Pydantic.

AI Engine: llama.cpp python server (OpenAI-compatible endpoint).

Infrastructure: Docker, Docker Bake, Caddy (Reverse Proxy).

⚡ Quick Start

Clone & Configure:

git clone https://github.com/vlimkv/ai-project-tracker.git
cp .env.example .env

Launch Ecosystem:

# This will download the LLM model (~2GB) on first run
docker compose up -d --build

Access Points:

Web UI: http://localhost:3000

API Docs: http://localhost:8000/docs

LLM Stream: http://localhost:8080/v1

⚙️ Environment Variables

The system is highly configurable via .env. Key parameters:

AI_PROVIDER: Set to oss for local llama.cpp or openai for cloud fallback.

OSS_MODEL: Specify the HuggingFace model tag (default: qwen2.5-3b-instruct).

REDIS_URL: For FSM state storage and AI response caching.

⚠️ Requirements

Docker Desktop installed.

RAM: Minimum 4GB (8GB recommended for larger models).

GPU: Optional (supports CPU inference via llama.cpp).

---

Developed by Alimkhan Slambek. Architecture demonstrates secure, private AI integration pattern.