ProVision

Created by team MeowVision on July 09, 2026
Video Captioning

ProVision is a video captioning agent built for Track 2. It analyzes short video clips and returns captions that match the requested tone, including formal, sarcastic, humorous technical, and humorous non-technical styles. The system focuses on describing visible subjects, actions, setting, and mood clearly while avoiding unnecessary assumptions. It is designed to generalize to unseen clips, produce concise JSON output for automated judging, and stay within the Docker-based evaluation workflow. For each task, ProVision samples representative visual moments from the video, interprets what is happening, and generates polished style-specific captions that preserve factual accuracy while matching the target writing style. The goal is simple: reliable captions that are useful, readable, and aligned with the evaluator format.

Category tags:
Explore more applications

ARouter

A highly optimized, resource-aware routing engine designed to process complex AI queries under strict hardware constraints. It intelligently balances local compute and cloud APIs to maximize accuracy while preventing out-of-memory failures.

TheWho

LLaMA

ForgeMind AI–Intelligent Software Architecture

ForgeMind AI is an AI-powered software architecture assistant that transforms project ideas into complete technical blueprints, including system architecture, security recommendations, development roadmaps, and implementation guidance and etc.

Neo Dev

AI StudioAI/ML APIChatGPTClaude CodeCodexGemma 2GemmaGemini AIGemini 3 proGemini 3 FlashGenerative AgentsGenerative AI StudioGithub CopilotOpenAI

SEV-Cap: Gemma powered video captioning

SEV-Cap captions short videos in four styles—formal, sarcastic, humorous-tech, and humorous-non-tech—using Gemma end-to-end. Facts are verified with semantic-entropy before captioning; every caption must pass a blind style lineup and self-refine loop.

sakku

GemmaCursor

Zero-Token Routing Agent

A local-first routing agent for Track 1. It solves math, logic, and code locally with a small Gemma model and verified execution, and answers language tasks on the spot. The goal is zero Fireworks tokens while staying above the accuracy gate.

Tensen

AMD ROCmGemmaHuggingFace HubLlama 3.1

asdaaaaaaaaaaaaaaaa

asdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaaasdaaaaaaaaaaaaaaaa

huSAM

AMD Developer CloudClaude Code