Over the past few months, I’ve been building Thothium, a local-first desktop video studio designed to automate the "faceless content pipeline" end-to-end.
Most tools in this space are black-box SaaS wrappers where you submit a prompt and hope for the best. I wanted something that does the heavy lifting but keeps a fully interactive, scene-first timeline Non-Linear Editor (NLE) in the loop so you actually have creative control.
Give it an abstract topic and the pipeline handles:
The Technical Part I’d Love Feedback On:
Everything in the architecture is content-addressed (SQLite queue + BLAKE3 cache store).
Because of this, if you change a single sentence or swap an image in the middle of your timeline, the app doesn't re-render the whole video It performs a "smart render" - re-voicing and re-encoding *only* the specific scene you touched, while stream-copying the rest of the timeline cache into the final export container. You get an exact, lightning-fast draft preview rendered through the real pipeline.
The app engine is written end-to-end in Zig, utilizing a native webview shell (Svelte 5 + Tailwind) for the editor interface, and it shells out to a persistent Python worker to keep the 1.7B TTS model hot in VRAM.
Current State & Limitations
Right now, this is a desktop app tailored for a Windows + NVIDIA setup (the TTS wants \~4GB VRAM and whisper takes \~1.5GB). I’m planning out CPU fallback options for the TTS engine, alongside macOS/Linux compilation shims down the road.
Any changes you think I should make with my current architecture?
I am looking for feedback from content creators and developers on the overall workflow:
What visual effects, transitions, or timeline controls are missing for your use case?
How does the text-to-timeline editing pacing feel?
Any changes you think I should make with my current architecture?
(Note on the attached video: The project overview dashboard at the start is missing a few thumbnails because the frontend is brand new and I'm still backfilling the cache registries! Also, I originally sped up these clips thinking I needed to squeeze them into a strict .gif limit before realizing how well Reddit handles raw MP4 looping now.)
If you're interested in local-first AI pipelines and want to poke around a raw native binary, let me know what features you'd want to see plan on rolling out an alpha phase over the next month, and then proceed into a paid beta the following 1-2 months as I move it into a SaaS format.
Thothium
https://reddit.com/link/1upmyj5/video/69td7ruk5rbh1/player