Video Captioner

Video Captioner is a 5-stage AI pipeline for accurate, multi-style video captions built for the AMD Developer Hackathon Track 2. The Problem: Speech-to-text systems make mistakes. Whisper transcribes "Ollama" as "Olimar" — basic pipelines pass these errors to captions. Our Solution: A 5-stage validation pipeline that catches and corrects errors using context: Stage 1: Dual Whisper Transcription — runs Whisper twice with different temperatures to capture variation. Stage 2: Theme Detection + Cross-Comparison — single LLM call detects video topic and resolves disagreements using theme context (not just "most accurate"). Stage 3: Quality Check — fixes errors without rephrasing correct content. Stage 4: Caption Generation — produces 4 styles (Formal, Sarcastic, Humorous-Tech, Humorous-NonTech) while preserving ALL information from the master caption. Stage 5: Emotion & Tone Verification — scores each caption 1-10 and auto-regenerates if below 8. Results: 100% success rate on 6 test videos, 66 seconds average processing time, context-aware error correction (Olimar→Ollama), and full content preservation (140-154% of master caption length). Technology: Python, Whisper (local), Gemma 4 31B via Fireworks AI, Docker containerized.