Build a narrow, single-purpose agentic tool that reliably replaces one real task, with the model generating responses running entirely on your own hardware.

• Local inference only: no hosted/frontier model APIs (OpenAI, Anthropic, Gemini, etc.) in the final submission
• Open-source or open-weight model
• Non-generative external tools (web search, APIs) are allowed
• One clearly defined task, not a general assistant
• Public GitHub repo, open-source license
• State the model used, its parameter count, and minimum hardware tested on
• Show 3 repeated runs demonstrating consistent output
• Scored on task success rate, reliability, and practical usefulness, with tokens-per-second as a tie-breaker