Track 2, Harish K, Decompile

[harishkotra]

Decompile

Track: Track 2 - Development & Local Deployment of Private AI Agents

Overview

Decompile is a local-first autonomous software engineering system that transforms YouTube videos into production-ready software project packages. Instead of producing only a summary, it creates requirements, architecture, API contracts, frontend and backend stubs, Docker assets, deployment guidance, QA plans, documentation, and a downloadable repository ZIP.

AMD Radeon usage

• Local LLM runtime adapter boundary for vLLM ROCm or llama.cpp HIP
• Local Qwen2.5-VL vision and PaddleOCR adapter boundaries
• Local BGE/FAISS-ready memory and embeddings
• ROCm/HIP telemetry contract with GPU model, VRAM, utilization, tokens/sec, latency, embedding speed, and vision timing
• Reproducible local benchmark method and baseline

Repository contents

• Complete Next.js and FastAPI source code
• Specification PDF and PPT/PDF slides
• Benchmark, architecture, API, deployment, local-AI, and test documentation
• Docker Compose configuration and CI workflow

Thank you for reviewing this submission!