CaptionForge-THAI

CaptionForge is our Track 2 submission: an AI agent that watches a video clip and generates one caption per requested style (formal, sarcastic, humorous-tech, humorous-non-tech), built entirely on Fireworks-hosted models with no other API dependencies. Instead of the usual multi-stage pipeline — a describer, a writer, and a judge passing the clip along — CaptionForge uses a "direct and guarded" design: one vision model (qwen3p7-plus) looks at four sampled frames and the style brief together and writes the finished caption in a single call, once per style, in parallel. A different model family (kimi-k2p7-code) stands by as a spare, used only if the primary is still unreachable after retries. Cutting the relay means less latency and fewer failure points — with no LLM judge involved, style rules are enforced by deterministic code instead: regex catches emoji, banned slang, a missing required tech term, or a stray exclamation mark in formal, and a caption that still breaks a rule gets exactly one regeneration naming the exact problem. Accuracy guardrails are built into the prompts themselves: the model is instructed never to quote or transcribe on-screen text (compressed video makes OCR unreliable, and a misread sign costs more than a generic description) and never to name a real-world city, country, or landmark. The last layer is reliability. Clips are probed and processed heaviest-first against a hard wall-clock budget. If a caption call fails on a transient API or network blip, it isn't marked as a permanent failure — leftover time budget funds a paced salvage pass that retries it once the blip clears, so a temporary hiccup doesn't cost real accuracy points. Only if every attempt is exhausted does a style fall back to a tone-correct generic caption — never a blank one. Whatever happens, the container always exits 0 and writes a valid, fully-filled results.json within the 10-minute limit.