Nimbus AI

Nimbus AI is a state-of-the-art, 100% on-device video captioning and copywriting cockpit that executes entirely in the user’s web browser with zero server cost, zero cloud latency, and complete data privacy. Powered natively by on-device WebGPU acceleration and the Gemma 4 E2B IT multimodal model, Nimbus translates visual storytelling into diverse literary and marketing styles locally on the client's hardware. Our architecture features a highly optimized Two-Stage Generation Pipeline: - Stage 1 (Factual Grounding): Performs synchronized, non-blocking temporal frame extraction on the browser thread, then compiles these frames into high-precision, normalized tensors. These are passed to Gemma 4 E2B IT for an objective, multi-layered chronological scene description. - Stage 2 (Stylistic Copywriting): Discards heavy visual tensors to preserve VRAM, utilizing Gemma's text-only engine to sequentially synthesize distinct captions (Formal, Sarcastic, Humorous Tech, and Humorous Non-Tech). Signature Technical Breakthroughs: - Bypassing the Transformers.js VLM Bug: Overcomes the library's broken pipeline('image-to-text') preprocessor bug by manually orchestrating AutoProcessor and raw .generate() inputs. - AMD Power Governor: Features active hardware throttling (Turbo vs Eco) that scales VRAM limit (384px vs 224px) and sampling logic to regulate system thermal footprints. - Near-Native Caching: Uses Origin Private File System (OPFS) for sandboxed NVMe model-loading read speeds.