First Principles Manufacturing, a newsletter by Yuanwei Ma at the intersection of AI, robotics, manufacturing and US–China.
What this is
Dispatches from a robotics lab in Novi, Michigan, on what AI and robots actually look like on a factory floor — inspection, grinding, deburring, welding — and why so much of the work fails to translate from PowerPoint to production.
Who I am
I’m Yuanwei. PhD in nuclear engineering from Shanghai Jiao Tong University — radiation protection and radiation measurement, where the foundational skill is pulling a real signal out of background noise in a way that’s auditable and reproducible. In 2020 I moved that methodology into industrial vision.
The last six years, on factory floors in China:
Led the development of a Flexible Industrial Surface Defect Inspection System that generated roughly $28M USD in commercial revenue, deployed at Microsoft (PO #100667750), Apple, Amazon, BYD, JAC Motors, Jinyan Technology, and China Mobile.
Since 2024, led an Embodied Intelligent Industrial Robot platform.
Built three separate attempts at productizing the stack — not another one-off project, but something repeatable across customers.
Along the way: about 50 granted patents in machine vision and robotics; co-authorship of two Chinese national AI standards (GB/T 42755-2023 Technical Requirements for Machine Learning Systems and GB/T 43782-2024 Data Annotation Procedures for Machine Learning), with ongoing participation in an IEEE standard currently in drafting; a paper at AAAI 2025 (the 3CAD dataset for industrial anomaly detection); and the Wu Wenjun AI Science and Technology Award as 1st ranked contributor — China’s most prestigious AI prize, functionally analogous to the ACM Prize in Computing.
I currently serve as Chief Scientist, overseeing R&D direction and major technical decisions at Changzhou Microintelligence (~400 people, pre-IPO) and lead its U.S. subsidiary, Micro Intelligence Inc., in Novi as Head of Technology R&D Center.
That’s the CV portion — what slides up on the deck. The part I care more about is what never makes it onto a deck: the 1 a.m. customer calls, the plant manager’s death stare, the blame loop between equipment vendors and algorithm teams, and the tired look an AI engineer gets when yet another new defect class shows up.
I didn’t learn manufacturing from papers. I learned it from standing next to the machines.
Why this Substack exists
Manufacturing AI, as a public narrative, is dominated by two things:
Fundraising decks — which need everything to sound inevitable.
Vendor marketing — which needs every product to sound SOTA.
The shared problem: neither tells you what the engineering actually looks like.
I’ve watched a lot of factories “adopt AI” where nothing changed on the floor — a big screen on the wall, cameras installed, a polished deck, and stacks of if-else patches underneath. Two years later no one on the team will touch it.
This Substack is for the other half of the story: the real engineering, the real numbers, the mistakes that actually cost money.
If you work in factory automation, run an industrial AI lab, make decisions about a supply chain — or just wonder what’s happening behind all those cameras — welcome.
What I write about
Four content pillars:
Deep Dive — long-form, 2,000–4,000 words. Specific cases, first-principles breakdowns, physics and math on the page.
Weekly Intel — shop-floor data and short commentary. A specific number, a project fragment, a moment from the floor.
US–China Compare — how manufacturing differs across the two, told through engineering and physics, not geopolitics.
Pitfall Guide — inversion thinking. How to lose money building manufacturing AI. Usually more useful than how to make it.
Plus Short-form Notes — lab photos, data screenshots, a single observation. The unpolished side.
What I don’t write about
No courses for sale.
No “smart manufacturing transformation” consulting pitches.
No “general AI will remake everything” takes.
No vendor advocacy dressed up as analysis.
No “Company X raised $100M” news coverage.
A note on disclosures
Some of the projects I reference were delivered through my current company, Micro Intelligence Inc., or its parent, Changzhou Microintelligence. I’ll flag it when that’s the case. Opinions here are my own and don’t represent my employer. This Substack is not a channel for product pitches or client inquiries — those go elsewhere. Customer-specific material appears only with appropriate authorization or anonymization.
Who this is for
Engineers working on factory automation.
Researchers in industrial AI labs.
Operations and supply-chain decision-makers.
Investors covering manufacturing.
Anyone who likes their engineering with the physics still visible.
If you’re not in that list, you’re still welcome — but the density is calibrated for the people above.
A stance I’ll keep coming back to
Iron Man suit, not lights-out factory.
AI isn’t replacing factory workers — it can’t, and the reasons are physical, not cultural. AI’s job is to amplify them: let a trained QC operator go from reviewing 1,000 parts a day to the 50 high-risk ones; let a welder hit sub-millimeter consistency with the right feedback stack behind the torch.
Most fundraising decks sell “fully autonomous, 24/7, zero humans.” What I’ll defend here is a different path — human-in-the-loop, tight data feedback, every layer with clear capability boundaries. Slower. Less sexy. And the only one that actually runs on today’s AI.
Cadence
Long-form: one piece per week, rotating through the pillars.
Notes: daily, five minutes to read.
Everything free for now — a paid tier may come later; not yet.
Technical debate, collaboration inquiries, well-reasoned disagreement — all welcome. Please don’t email asking me to “turn around our stalled AI project.” That’s a paid conversation, and it isn’t what this Substack is for.
If you want to see what the engineering actually looks like, please:
