> loading manifesto...
We believe video intelligence
should run on your machine.
Not on someone else's server. Not behind a login wall. Not metered by the minute. The silicon in your PC is the most powerful creative tool ever built — and the entire SaaS industry exists to make you forget that.
[01] The Problem
Every year, video editing tools get more powerful — and more dependent on infrastructure you don't own. Your footage travels to a data center. A model on a GPU farm analyzes it. The result is streamed back. You pay per minute, per render, per export.
The pitch is always the same: "Let us handle the hard stuff." But the hard stuff is already sitting in your PC. Your hardware runs the same CLIP model that OpenAI charges $0.02/image for. That NVMe drive? It reads 7 GB/s — faster than any cloud download. The only thing missing is software that respects the hardware you already paid for.
[02] Our Answer
Onset Engine is a kinetic orchestration engine. It analyzes your footage using 768-dimensional CLIP vectors, maps your music track into beats, energy curves, and drop zones, then assembles a mathematically optimal edit — all locally. No files leave your machine. No internet connection required after activation.
OpenCLIP ViT-L/14 running locally — GPU-accelerated when available, fully functional on CPU. Every frame gets a 768-dim embedding. The engine doesn't guess — it calculates.
librosa decomposes your track into beats, onsets, and energy envelopes. Silent tails are auto-trimmed. Drops are detected. Cuts snap to musical precision within ±200ms.
JSON files define what "intensity" means for your genre. Contrastive scoring ensures tier specificity — not just raw similarity. Content-agnostic by design.
Chunked, memory-safe rendering with hardware encoding when available (NVENC, AMF). 4K footage on standard hardware. No swapping. No crashes. No cloud queue.
[03] What We Believe
Software shouldn't be a landlord.
Onset Engine is a one-time purchase. You own the binary. No monthly fees, no watermark extortion, no "upgrade to keep exporting." Pay once, render forever.
Your footage is not training data.
We never see your files. The AI runs locally. There is no telemetry, no analytics, no "anonymous usage data." The app phones home exactly once — for license activation — and never again.
Your hardware is underutilized.
Your CPU and GPU aren't just for gaming. They run the same transformer architectures that power billion-dollar AI services. We just cut out the middleman.
Composability beats features.
Presets are JSON. Drivers are JSON. The database is SQLite. The timeline exports to OpenTimelineIO. Every piece of Onset Engine is designed to plug into your existing workflow, not replace it.
[04] The Technical Stance
We don't compete with VC-funded cloud platforms on general-purpose editing. We stay in our moat: local-first CLIP vector search + beat-synced compositing. That's where the math is defensible, the user experience is differentiated, and the product can be maintained by a small team building on SQLite and FFmpeg — not Kubernetes and Stripe metering.
OpenCLIP ViT-L/14 Vision embeddings librosa Beat detection + energy mapping MoviePy + FFmpeg Chunked rendering NVENC Hardware encoding MPV Zero-latency DJ playback SQLite Local vector database Decord Fast frame sampling OpenTimelineIO NLE interop Ready to Take Control?
One engine. One purchase. Your machine. Your footage. Your rules.