VIDEO CTRL-F
Team consisting of EPITA Master's students (incl. Sorbonne M2), skilled in NLP/LLMs, Python/Java/C++, distributed systems and Mistral AI—hackathon-tested (Minecraft bot) and organized 1,000+ attendee events.
Project Description
Built a two‑service “Video Ctrl‑F”: Gemma3 (via Ollama) on Cloud Run GPU (NVIDIA L4) + a FastAPI/ADK agent orchestrating YOLOv8 object detection and Whisper transcription.
Core: LLM container deployed to Cloud Run GPU; agent container deployed and calling the LLM; reproducible with Docker and env‑based config.
Innovation/UX: fuses CV+ASR+LLM to return precise time ranges; simple POST /chat API and lightweight Streamlit demo; one‑click deploy + cURL walkthrough.
Stack: Cloud Run (GPU L4), Docker, Python, FastAPI, Google ADK, LiteLLM, Ollama, Gemma3, YOLOv8, Whisper, PyTorch, OpenCV, NumPy, ffmpeg, yt‑dlp; scaling via agent concurrency, LLM min instances, timeouts/retries, structured logging