Multi-Agent-Hackathon-Suite

Hackathon Agents is a collection of three independent AI agents designed for practical, efficient AI workflows. 1. General-Purpose Agent: A cost-efficient task router that handles all required task categories using a local model by default. It automatically infers the task category from the prompt using lightweight lexical rules, avoiding model-based classification. Only when a local response fails a validity check does it escalate to the Fireworks API, minimizing paid token usage while maintaining response quality. 2. ROCm Ready: An AI-assisted CUDA-to-ROCm migration tool for AMD GPUs. It scans GitHub repositories or code snippets to detect CUDA-specific components such as kernels, runtime APIs, GPU libraries, and build configurations. Using AMD's official hipify-perl tool, it performs mechanical code conversion where possible and categorizes findings into Auto, Install-only, and Manual migration steps. A local LLM (Ollama with qwen2.5-coder:3b) generates a clear, human-readable migration report based solely on verified tool outputs. 3. Video Captioning Agent: A fully local video captioning pipeline that generates captions in four styles—formal, sarcastic, humorous_tech, and humorous_non_tech—for every input clip. The solution requires no external APIs, API keys, or cloud services, ensuring offline execution, zero usage costs, and compatibility across regions while meeting the hackathon's deployment requirements.