ChronosCap: Multi-Modal Semantic Video Captioner

ChronosCap is a high-performance, multi-modal AI agent built to generate styled captions for diverse video content. The system processes videos using a four-stage pipeline: ingestion, multi-modal perception, semantic contract resolution, and single-pass styled generation with cross-style validation. First, the Video Loader downscales video frames to optimize token usage. The Adaptive Sampler dynamically scales keyframe sampling density based on visual complexity and scene boundary transitions, ensuring that high-motion events are captured while static frames are skipped. Second, the Perception Layer runs parallel VLM frame description, OCR text extraction, and Whisper audio transcription. In the Groq implementation, all keyframes are bundled into a single multi-image VLM request to reduce API call overhead and prevent rate limits. Third, the Semantic Contract Resolver fuses timeline data into a unified, chronologically-consistent, style-neutral narrative. This central narrative serves as a factual anchor, preventing the hallucinations commonly associated with raw captioning. Fourth, the Single-Pass Generator rewrites the narrative into four distinct captions (formal, sarcastic, humorous_tech, and humorous_non_tech) in a single LLM completion. Word budgets are enforced via deterministic sentence-boundary trimming and LLM editor fallback rewriting. Finally, the Semantic Validator performs a consolidated cross-style check. It verifies factual adherence, measures lexical Jaccard distances to guarantee style separation, and outputs an Evidence Justification Record. If API issues arise, a local rules-based solver acts as a fail-safe, ensuring complete results under the strict 10-minute runtime budget.