.png&w=256&q=75)
🤓 Latest Submissions
TRANSDUCER — Style-Aware Video Captioning Agent
TRANSDUCER is an autonomous Track-2 agent that watches a video and retells it in four distinct voices: formal, sarcastic, humorous_tech, and humorous_non_tech. Instead of brittle rule-lists, each style is a character — a precise observer, a world-weary wit, a burnt-out engineer, a bewildered dad — handed the actual video frames so the humor reacts to what is genuinely on screen and never invents details the LLM-judge would penalize. Per clip (~15–25s): the download is streamed and salvaged even when truncated, four keyframes are sampled across the whole clip at 1024px, and a single multimodal call to Kimi K2.6 on Fireworks AI returns all four captions at once — forcing sharp contrast between voices — followed by per-style retries, a deterministic fallback ladder guaranteeing complete valid output, and an artifact sanitizer. Engineered for the real grading box (4GB RAM, 2 vCPU, keyless, cold anonymous pull): three clips run concurrently, every clip finishes under 30 seconds, and no style is ever missing. Across 20+ leaderboard-scored builds — one change at a time, kept or reverted on the number — it climbed 0.61 → 0.84 → 0.91, peaking at 0.92, top-3 on the hidden set.
13 Jul 2026
