StyleForge — Four-Style Video Captioner

StyleForge captions video clips in four styles (formal, sarcastic, humorous_tech, humorous_non_tech) with a three-stage pipeline: Kimi K2.6 vision (via Fireworks AI) watches the clip and writes a dense factual description; a Gemma 3 4B — DPO-fine-tuned on AMD Instinct MI300X and quantized to run on CPU inside the container — writes candidate captions; an LLM judge reranks candidates per style. The RL story: we built a 1,200-cell training factory (300 synthetic scenes × 4 styles × 6 candidates, all judge-scored), then ran two measured DPO rounds on a single MI300X (~8 minutes each). Round 1 lifted sarcastic tone +0.54 over base; after inspecting failures we rebuilt preference pairs accuracy-weighted, and round 2 lifted formal +0.50 — bringing the 4B to exact parity with the ~1T-class frontier model on factual accuracy (7.62 vs 7.62 on our independent-judge eval). We ship the mode that measured best (best-of-N) and the tuned Gemma rides in-container as a zero-dependency fallback voice. Engineering for the 10-minute wall: a schema-valid results.json is written at startup and atomically upgraded per task; three-tier degrade (best-of-N → single-shot → fallback captions) means dead URLs, API outages, or malformed tasks never zero the run. The 3 example clips complete in 33 seconds. Built by a two-member team (Sankar Subbayya × Claude) on AMD Developer Cloud (ROCm 7.2.4), Fireworks AI, and Google DeepMind Gemma.