Local on-device AI versus cloud services

In this episode of Crazy Wisdom, Stewart Alsop sits down with Andre Oliveira, founder of Splash N Color, a bootstrapped 3D printing e-commerce business selling consumer goods on Amazon. The two cover a lot of ground — from how Andre went from running 40 FDM printers out of South Florida to offshoring manufacturing to China, to how he's using Claude Code to automate inventory management and generate supplier RFQs across 200+ SKUs. The conversation stretches into bigger territory too: the San Francisco AI scene, the rise of AI agents and what they mean for the future of the internet, whether local on-device AI will eventually replace cloud-based tools, and why building physical products will stay hard long after software becomes easy. It's a candid, wide-ranging conversation between two self-taught builders figuring things out in real time. Follow Andre on X: @AndreBaach.

Timestamps

00:00 — Andre introduces Splash N Color, his Amazon-based 3D printing e-commerce business and explains the grind of running 40 FDM machines in South Florida.

05:00 — The conversation shifts to Claude Code and how Andre built an inventory automation system to manage sales velocity and RFQs across 200+ SKUs.

10:00 — Stewart and Andre compare notes on Opus 4.6, debate Codex vs Claude, and Andre breaks down the new Agent Teams feature in Claude Code.

15:00 — Discussion turns to the San Francisco AI scene, the viral OpenClaw launch event that drew 700 people, and what's capturing the city's imagination right now.

20:00 — The pair wrestle with data privacy, the illusion of it since 2000, and whether full transparency of personal data might actually serve people better.

25:00 — Stewart pitches his vision of local on-device AI replacing cloud tools entirely, and they debate the 10–15 year timeline for mainstream societal adoption.

30:00 — Andre traces his origin story: a high school dropout from Brazil who spotted a 3D printing opportunity on Facebook Marketplace and got lucky timing with COVID.

35:00 — They explore whether AI-generated 3D models and DfAM will automate physical manufacturing, and why proprietary specs keep the space stubbornly hard.

Key Insights

1. Lifestyle businesses deserve more respect. Andre spent months feeling inadequate scrolling through Twitter watching founders announce funding rounds, before realizing his cash-flowing, location-independent business was already the goal. The social media version of entrepreneurial success warped his perception of what he actually had built.
2. Claude Code is becoming an operating system. Stewart describes running Claude Code as having a second OS on top of MacOS — one that makes the underlying machine legible in ways it never was before. Both guests use it not just for coding but as a primary interface for understanding and operating their businesses.
3. Agent Teams changes how work gets done. Andre explains that Claude's new multi-agent feature lets you assign a team lead and specialized roles that communicate with each other in parallel, essentially running an autonomous task force inside your terminal — a meaningful leap beyond single-instance prompting.
4. Physical manufacturing will stay hard. Even as AI-generated 3D models improve, tolerances of 0.5 millimeters can mean the difference between a product working or not. Design for manufacturing is a separate discipline from design itself, and proprietary specs mean open source models rarely hit commercial quality.
5. The internet is heading toward agents. Both guests agree that AI agents will increasingly handle tasks humans currently do manually online — booking services, making payments, coordinating logistics — with the human internet potentially becoming secondary to a machine-to-machine layer.
6. Iteration is the real value of 3D printing. Andre pushes back on 3D printing as a business unto itself, framing it instead as a prototyping tool. The true value is rapid iteration on housing, tolerances, and fit — not the printer, but the speed of the feedback loop it enables.
7. Technology compounds in layers. Andre closes with a tech-tree analogy: each generation normalizes the tools of the previous one and builds the next layer on top. Agentic coding today is what the internet was in the 90s — the foundation for something we can't yet fully see.