Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Swift + Jetson: Robots in Minutes
TL;DR
Wendy Labs is trying to make robot development feel like iPhone app development, not a three-day Linux ordeal — CTO Yanis says even MIT robotics grads still lose days flashing Jetsons and wrangling peripherals, so Wendy built a CLI, OS, and agent stack to get devices running in 10–20 minutes with remote debugging and OTA updates.
Their bet is Swift, not Python, as the bridge between prototyping and production robotics — Yanis argues most labs prototype in Python then rewrite in C++, while Swift gives them high-level ergonomics plus C/C++ interop, native performance, and memory/concurrency safety across everything from microcontrollers to Jetson-powered robot dogs.
The hardware range is surprisingly broad: $8 microcontrollers, $250 Jetson Nanos, $5,000 Jetson Thor boards, and even humanoids — Wendy says the same infrastructure can span cheap industrial sensors, Raspberry Pi 3/4/5, Jetson Orin modules, drones for forest-fire detection, autonomous RC cars, and heavy humanoids that need harnesses because they absolutely do not know not to step on your foot.
The most concrete demo was a Unitree-style robot dog that Wendy got following people in about two days, with only 10–20 minutes of setup using their tools — the dog runs a YOLO-based person-following stack with voice commands, LiDAR/light sensors, and RealSense-style depth sensing, and the team claims the polished version should become a ‘2-minute, two-command’ setup.
A big part of the pitch is retroactive AI on edge devices: deploy first, ask new questions later — Yanis describes using MCP, LLMs, and lightweight models to push new logic onto existing drones or factory sensors after the fact, like adding people-detection to a fire-spotting drone or mining microphone data in an oil facility for engine anomalies you didn’t think to monitor originally.
For AI builders, the real wedge is infrastructure speed, not flashy demos — Wendy is pre-seed with about eight developers, has OS CI builds down to 5 minutes, open-sourced the stack under Apache 2.0, and claims they reached in 2–3 months what one autonomous vehicle company spent 2–3 years and a lot of money building internally.
The Breakdown
Live from a San Francisco robot lab
Ray opens in Wendy Labs’ SF hardware space — robot dogs on the floor, a drone mid-build, humanoids nearby — with the exact hackathon pain point every robotics person knows: half your weekend disappears before the robot even boots your code. Yanis, Wendy’s CTO, immediately frames the company around that frustration: robotics is still weirdly harder than shipping your first iPhone app.
Why Jetsons are powerful — and miserable to set up
Yanis walks through the practical misery of edge robotics: flashing Nvidia Jetsons, hunting for HDMI adapters, keyboards, displays, and broken updates just to run a simple YOLO model. His contrast is blunt and memorable — you can buy an iPhone, plug it into a Mac, and get an app running in 10–20 minutes; with a Jetson, many people burn days just installing the OS.
Swift as the anti-rewrite strategy
The big technical thesis lands here: Swift gives Wendy a way to avoid the classic robotics workflow of “prototype in Python, then rewrite in C++ when it matters.” Yanis, who says he’s spent about 12 years with Swift and built Vapor and Hummingbird, pitches Swift as the rare language that is high-level and approachable but still low-level enough for native performance, C/C++ interop, and embedded use.
A robot dog that can follow you, plus a lab full of edge AI bodies
Ray gets the fun part: a Jetson-powered robot dog Wendy already reprogrammed to follow people autonomously, respond to voice commands like “sit,” and run local vision on-device. The team shows off stabilizing legs, spinning LiDAR/light sensors, a manual override remote, and the very real safety mindset behind the scenes — open test space, harnesses, and a repeated reminder that these machines are still “empty shells,” not movie robots.
From warehouse patrol to wildfire drones
Yanis shifts from demos to use cases: patrolling warehouses, detecting intrusions without sending all footage back to a human, and building drones that can spot early forest fires with infrared or other sensors. The especially interesting part is his “retroactive” AI idea — if a drone was only looking for fires but suddenly needs to identify nearby people, Wendy wants an LLM to push new logic or a small model onto the mission without months of re-plumbing telemetry and cloud systems.
The hardware ladder: from $250 Jetsons to $5,000 Thor boards
The camera moves to a table of Nvidia hardware, and the numbers get concrete. Yanis calls out an 8GB Jetson Nano at around $250 for YOLO and speech tasks, an AGX Orin with 64GB unified memory for heavier VLM-style workloads and autonomous vehicles, and a Jetson Thor with roughly 128GB unified memory at around $5,000 for high-bandwidth sensor stacks like LiDAR in cars.
The developer experience pitch: one cable, open source, debug anywhere
Wendy’s product story gets specific: a single-command CLI, USB-C development with no Wi‑Fi required, local debugging over Wi‑Fi, and cloud brokering over mTLS where the developer — not Wendy — controls the device. Yanis keeps returning to a dead-simple standard: if something that should be easy takes more than an hour, it’s either a docs issue, a missing template, or a bug.
A tiny pre-seed team trying to compress years into months
Near the end, Wendy sounds less like a robotics startup and more like an infrastructure company with robot-shaped proof points. They’re pre-seed, about eight developers, distributed across the Netherlands, Germany, Romania, Tasmania, the US, and Canada; they’ve pushed OS build times down to five minutes, are open source under Apache 2.0, and Yanis says one autonomous vehicle company spent 2–3 years building infrastructure Wendy matched in 2–3 months. The closing energy is basically: the demos are fun, but the real win is shaving every recurring five-minute frustration until robot iteration feels like modern software.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
