Summary
In this video I walk through my microfactory node – a simulated 3‑D‑printer interface that lets anyone load, slice, preview and print 3‑D models without needing an expert at their shoulder. I start by explaining why I built it: printing at home with my family is fun but often runs into hidden issues that only seasoned engineers can spot. The UI mirrors a standard slicer with four main sections – Load, Slice, Print and Review – and runs entirely in a simulated environment that you can explore via the README and notebook files. I show how you can position the part, choose materials, randomise settings, reset or even override the defaults.
The backend is built on a small “gym‑of‑four” model suite, originally the e4‑b model. I collected datasets, fine‑tuned the model and system prompts, and tested several variants. The LoRa‑v3‑q8 quantised model turned out to be the best, so I warm it up before use.
Next I demonstrate loading a part – the classic Benchy – and how every job is automatically logged to a ledger and stored in a dataset for later analysis. After loading, I slice the model; the slicer is a fast, custom implementation in Gradient that instantly generates layer geometry. You can scrub through the layers or replay a virtual print to see how the part will be built.
The real magic happens when the AI “engineers” take over. I introduce Chief Engineer O’Brien (a Star Trek reference) who reviews past jobs, checks the ledger and suggests a printing plan, flagging potential failure zones such as sagging at a bridge. A second opinion comes from Engineer LaForge, who confirms that there are no red flags. The system then validates the plan with a deterministic “spine” module.
You can still override settings or tweak environment variables before launching the print iterations. I run eight simulated iterations; the system predicts whether the print will succeed, when it might fail, and when it converges to the best quality. In this case the simulation shows a clean pass across all iterations, matching historical runs, despite an over‑cautious inspector that tried to halt the job.
The Review tab aggregates all session data: inputs, O’Brien’s recommendations, LaForge’s review, simulated runs, outcomes, and any new policies learned. If you have a physical printer attached you can push the final settings to the real machine, log the actual print, and the results are appended to the same ledger, enriching the dataset.
I also highlight the broader vision of the micro‑factory system: a suite of interoperable tools that individual makers or small teams can share, exchange parts, knowledge, and policies, effectively creating a collaborative network of 3‑D‑printing expertise.
Finally I invite viewers to subscribe for updates, try the demo on Olama or Modal via remote APIs, or run it locally. The project is fully open‑source, hosted as a Gradio space for quick demos, with all models pullable from Hugging Face or olama.com. The code, runbook and README let you spin up the entire stack on your own hardware. I close by encouraging questions and wishing everyone good luck.