Thanatan Budsri@ThanatanBS

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.

To participate in the hackathon, simply click Sign up with AMD. If you are not already a member of the AMD AI Developer Program (ADP), you will need to create an account. ADP members can access AI tutorials, experts, and community support. In addition, new members can claim a free month of DeepLearning.ai pro, $100 in AMD GPU cloud credits, and $50 in Fireworks AI API credits. All hackathon participants will receive on hackathon launch additional compute and API credits for their projects. Ready to build what's next on AMD? Sign up with AMD below to secure your spot.