.png&w=256&q=75)
🤓 Latest Submissions
Tetrlense ai: Style-Conditioned Video Captioning
Tetrlense ai ("Four Voices, One Vision ") is a state-of-the-art video captioning agent engineered for the AMD Developer Hackathon: ACT II. The agent generates contextually precise video descriptions dynamically adapted to four unique styled personas: Standard, Casual, Formal, and Sarcastic. Key Architectural Innovations: 1. Single-Pass Frame Slicing: Replaced traditional frame-by-frame loops with a unified ffmpeg select filtergraph, dropping frame extraction latency from over 10 seconds to under 0.8 seconds. 2. Double-Pass Grounding: Utilizes a dual-stage inference workflow. Pass 1 constructs a visually grounded scene understanding model, and Pass 2 maps this context to style-specific personas to eliminate stylistic drift. 3. Zero-Crash Safety Systems: Built with robust exception handlers, JSON schema verification, case-insensitive key parsing, and localized backup retries. If the agent detects an environment issue, it gracefully generates backup summaries to ensure a 100% success rate. 4. Adaptive Timeout Circuit Breaker: Evaluates runtime progress in real-time. If execution time reaches 360 seconds, the agent automatically throttles down to an ultra-fast single-pass mode to guarantee successful completion.
13 Jul 2026
