Prathamesh Patil
← All projects
prathameshrp@devlog:~/projects/project-raphael

~$ cat project-raphael.md

Project RAPHAEL

Real-time Agentic Desktop Focus Agent featuring an LLM Load Balancer, fast regex pre-filtering, and a KDE Plasma HUD.

PythonGenAIGroqFlaskSQLiteQML

RAPHAEL (Real-time Agentic Process & Human Attention Evaluation Loop) is a background Linux desktop agent and frosted-glass HUD overlay that classifies focus context using an LLM pipeline, calls out distractions, and enforces focus rules.

Key Technical Highlights

  • Multi-Provider LLM Load Balancer: Round-robin key rotation across Groq (Llama 3.1 8B), Gemini 1.5 Flash, DeepSeek, and OpenAI with automatic HTTP 429 rate limit failover.
  • Zero Latency Fast Path: Bypasses network calls for known developer tools (VS Code, terminal, docs) via regex pre-filtering, reducing classification latency to 0ms for 90%+ events.
  • Async Thread Pool: Offloads LLM requests to background threading.Thread worker pools with a thread-safe window_cache dictionary, eliminating Linux desktop shell UI stutters.
  • Web Telemetry Dashboard: Serves real-time focus timelines, top application distribution charts, and dynamic AI coaching insights at http://localhost:5757/dashboard.