Video Captioning Agent

A video captioning agent for AMD Hackathon Track 2. It downloads video clips from provided URLs, extracts midpoint frames via ffprobe/ffmpeg (4/6/8 frames adaptive to duration, resized to ≤1024px), and sends them to kimi-k2p6 (Fireworks API) for vision-based scene description. A separate set of kimi-k2p6 text calls then generates captions in four distinct styles: formal, sarcastic, humorous_tech, and humorous_non_tech, with the formal caption injected as a locked grounding caption to prevent hallucination across humor styles. A regex tech-word tripwire guards humorous_non_tech against leaked tech vocabulary, triggering one retry if detected. The pipeline handles varied video content (nature, urban, animals, people, sports, food, weather, technology), uses ThreadPoolExecutor for concurrent clip processing with 120s per-clip timeout, and outputs structured JSON with placeholder fallbacks for any failure. Built with Python 3.11, requests, and python-dotenv. Containerized via Docker ( python:3.11-slim + ffmpeg only)