LMC Tech
Published Android app
LODORi
A meditation app that writes and voices a personalised session in about 25 seconds. React Native, FastAPI, subscriptions, ads, safety checks, and a public release on Google Play.
Our work / Text-to-film pipeline
Movie Factory
A film pipeline that runs unattended on a self-hosted GPU. A language model writes the story and shot list, image models produce keyframes with characters that stay the same across scenes, a vision model rejects the bad ones, a video model adds motion and sound, a voice reads the narration, ffmpeg cuts it. You get a contact sheet after twenty minutes so a bad film is caught early.
~1h
for a nine-shot, 45-second film
3
rerolls per keyframe if the vision QA rejects a character
~20min
until a contact sheet lands in Slack for review
0
cloud GPU spend: everything runs on our own hardware
The brief
Keep the characters, lose the babysitting.
Short-form video is cheap to consume and expensive to make. The hard part of AI film is not one pretty frame, it is the same face in shot seven as in shot one, and a pipeline that survives a crash at 2am.
- Slack bot in, film on your phone out, via Syncthing.
- Character continuity: frozen style bible, byte-identical character canon, reference portraits fed through the image model, location plates, deterministic seeds.
- Vision-model QA scoring every keyframe against the character reference before any expensive motion work.
- Resumable state: an out-of-memory kill resumes from the last finished shot; one bad shot can be rerolled alone; recuts need no GPU.
- Quiet hours so it yields the GPU to other nightly jobs; measured benchmarks kept in the repo.
- Output feeds an auto-publisher that posts to TikTok, YouTube Shorts and Facebook on a schedule.
Under the bonnet
Self-hosted, end to end.
Qwen3 on Ollama for writing, Flux 2 for keyframes and reference portraits, Gemma 3 vision for QA, LTX 2.3 for motion with native synced audio, Piper for narration, ComfyUI as the model host and ffmpeg for the edit, all on an Ubuntu box with an RTX 3060. Systemd path units run the queue.
- Python
- Ollama / Qwen3
- Flux 2
- LTX 2.3
- Gemma 3 vision
- Piper TTS
- ComfyUI
- ffmpeg
- systemd
- Slack
What it shows
- Video production as a pipeline, with QA gates instead of hope.
- Serious self-hosted AI on modest hardware.
- The content engine behind our own social channels.
AI social media manager
Social Jen
Give it one idea, a product photo or a video and it drafts, schedules and publishes to nine social networks, writes graded Etsy listings and renders your 3D models. Nothing goes out until you approve it.
AI to 3D print pipeline
modelgen
Describe a part or drop in a picture and get a watertight, print-ready STL. Exact parametric CAD written by a language model, or organic meshes from AI, both validated before a slicer ever sees them.
Next step