SageOx is a well-funded, early-stage startup at the frontier of how AI is shaping the world. At SageOx, we believe the future of work is deeply human and collaborative. As agents take on execution, coordination, and routine work, people gain space for what only humans do well: creativity, judgment, relationships, the work of imagining, deciding, and creating together. Humans become not only more productive individually, but better at working with each other, leaning on their AI coworkers to carry out execution and raise the quality of how they collaborate.

We're looking for strategic hires - fearless, creative, independent thinkers - that will shape the future of the company, and how work is done in the modern world.

There's no trail yet. Pioneering and breaking trails is one of our company's core values. You're one of the first ten people cutting it.

The Job

We're breaking the trail toward a world where humans and AI agents work as one team — and increasingly, part of that world is physical hardware. The Ox Dot is our first device: a small, always-there presence that gives the hivemind a body in the room. It's real, it's shipping, it's the hardest and most unforgiving surface we build on — and it's one of several hardware bets we're weighing for what comes next.

We're looking for a full-stack engineer whose center of gravity is our hardware — the first hire dedicated to it. A large portion of your time will be on the Ox Dot today: making it stable, making it fast and safe to change, and building the machine that lets a small team (and a fleet of agents) develop hardware the way we already develop software. The rest is genuinely full-stack — you'll write the cloud services the device talks to, build the app and web experiences people actually use, and stitch the whole thing together, from firmware in the field to the pixels on the screen. And we move fast: in a few months the Dot may not be the thing we're building at all. You're someone who flexes with that — bringing up a device from another manufacturer, or picking up a different corner of the stack entirely, without missing a beat.

The problem you'll own

Building on the Dot is genuinely hard, and we leap in anyway — that's the job. But the hard part can swallow the day: a change lands, something unrelated breaks, and hours disappear into a bisect, because nothing is exercising the Dot deeply and around the clock to catch a regression the moment it happens. The result is that we don't innovate nearly as much as we could — the challenge of debugging keeps stealing the oxygen.

Your job is to change that. You'll stand up a hardware test lab — a small fleet of real devices under automated QA around the clock — so a change is either green or it isn't, and a regression surfaces the moment it lands, not days later. When debugging stops eating the day, the whole team gets to pour that energy back into building the next thing.

And the destination is big: we're building toward fleets of millions of customer devices in the field. You can't run at millions when every change risks a days-long debug — the small fleet on the bench is how we earn the confidence to get there.

What you'll do

Drive firmware and hardware development. Today that's the Ox Dot — a small audio device: boot, memory, audio pipeline, connectivity, the whole stack that has to Just Work in someone's home or office. Tomorrow it's whatever device we choose next.
Build the development harness. The single highest-leverage thing here isn't a feature — it's the tooling that lets an agent drive a real device end-to-end: flash, exercise, observe, assert, recover, repeat. You'll extend and reinvent the APIs that make a physical device agent-operable.
Stand up 24×7 automated hardware testing. A small fleet of real devices under continuous, unattended test. Behavior-Driven Development (BDD) scenarios as the source of truth for "does the device actually do the thing," running on silicon, not just in a simulator.
Run the agentic software factory. Get one agent operating one device with zero humans in the loop — then a fleet of agents developing, exercising, and self-triaging the bench devices in parallel, filing what they find. This is the factory: it produces software and confidence, not units on a line — and it's how a small team keeps a fleet of millions healthy.
Own OTA and the fleet. Ship firmware safely to devices already in customers' hands — staged rollout, health-gated updates, and A/B rollback so a bad update never bricks a device. At millions of devices in the field, the update pipeline is the product.
Write cloud services and user-facing experiences. The device is only half of it — you'll build the cloud services it talks to and the app and web experiences people actually use (memory and agent systems, auth, real-time streaming), and keep those seams honest against our digital twins.
Get devices ready to ship. The pipeline that turns a bare board into a customer-ready unit: provisioning, first-boot, pairing and auth, and everything that has to be correct before a device leaves for a customer.
Evaluate what comes next. The Ox Dot is one of several hardware directions we're weighing. You'll be the person who brings a candidate device home, tears it down, and tells us what's real.
Own the physical product with our partners. Work directly with suppliers on the supply chain, and drive customizations to the outer shell and accessories — the parts a customer actually touches.

The bet we're making

The agentic feedback loops that transformed software are just now arriving for hardware — and we're building them. Coming at the Dot agent-first pushed us to invent new ways to develop on top of a physical device: APIs that let an agent flash, drive, and observe a real board; twins that stand in for the cloud and the TV; harnesses that bisect a bug in minutes. That frontier is wide open, and it's the fun part of the job.

So the bar here isn't how many lines of firmware you can hand-write — it's how well you wield agents. A 2026 AI-native engineer spends less time reading every line and more time building the harnesses, loops, and specs that produce the code, bringing the judgment, taste, and direction while agents do the typing. If that's already how you work, you'll feel right at home.

What we've already pioneered — and where you take it

In just a few months on the Dot we've built things most hardware teams don't have, because we came at it agent-first. Building the next generation of them is the job — not a side quest. Here's the frontier you'd be standing on:

APIs that make a physical device fully agent-operable. HTTP and MCP surfaces let an agent flash the device, drive the touchscreen — tap, hold, swipe — pull a screenshot of the live screen, tail logs, inject a known audio file as the microphone, and assert on real on-device state. End to end, no human hands on the board.
Agents that look at the screen and reason about what to test. An agent captures a screenshot, reasons about what's actually rendered, decides the next thing worth exercising, and walks the UI like a person would — then cross-checks what it saw against ground-truth device state, so "the screen looks right" is never confused with "the device did the thing."
Digital twins of everything the device talks to. The Dot casts to TVs across four protocols (Chromecast, AirPlay, Roku, DIAL), streams audio to the cloud over MoQ, and talks to our API. We built faithful twins of all of it — conformance oracles that grade every message ok / warn / violation — so an agent can prove the full end-to-end path without a single real TV, relay, or cloud call.
A diagnostic harness that bisects by dependency layer, not by git history. Boot-time test sockets slice a silent hardware bug down the stack (mic → audio → wifi → net → auth) and converge in minutes on problems that used to eat days.
Micro-optimization as a daily sport. The entire product — audio pipeline, TLS, an animated round LVGL UI, WiFi, OTA with A/B rollback — lives in 512 KB of internal SRAM and 8 MB of PSRAM, where a single TLS handshake alone wants ~40 KB contiguous. Packing a rich device into that envelope, and keeping it there change after change, is real, constant, satisfying craft here.

This is the frontier. Widening it is the job.

Who you are

You use agentic AI coding as a daily practice, not an experiment. Claude Code (or equivalent) is how you already work, on hardware and everything else.
You build the system that builds the code. Harnesses, loops, specs, guardrails — you bring the judgment and direction, and you're at ease reviewing what an agent wrote rather than typing every line yourself.
You've built on real hardware. A shipped consumer product, or the firmware for a device you hacked together on your bench — either counts. What matters is you've built some complex code that runs on a physical thing, not the logo on the box.
You're a tinkerer and a hardware hacker. You reach for a logic analyzer and a soldering iron, and you think outside the box about how to make a device controllable, observable, and abusable by automation.
You think in systems, not functions. You see the whole loop — device, harness, cloud, agent — and you build the thing that makes the loop tighten.
You build and ship on your own. Side projects, finished and in the world. Active open-source contribution is a strong signal.
You're willing to bet on unproven approaches when the proven one is too slow.
You have agency and drive. You see what the product needs and go make it happen — nobody hands you a spec. That's the bar for everyone here, junior or senior.
Bonus, not required: you've led before. Directing a fleet of AI coworkers through a hard problem draws on the same muscle as directing a team through one — decompose, delegate with intent, review what comes back, know when to step in and do it yourself.

You might not be a fit if

We'll say this plainly, because it's the crux of the role — this job assumes an AI-native way of working. You may not be a fit if:

Your instinct is to hand-drive every test on the bench yourself, and the idea of an agent doing it makes you uneasy.
You want to write firmware and consider tooling, harnesses, and test infrastructure "not real engineering" or someone else's job.
AI coding tools are something you've tried, not something you live in.

None of that is a knock on great engineers. It's just not this job.

How We Hire

We don't run technical interviews. No whiteboard algorithms, no four-hour system-design rituals. Clear an initial assessment and we'll invite you to a work trial: 3–5 days, full-time, in person, building something real alongside the team, the way the team actually works. The trial will be on terms that are mutually agreed upon. Most of what either side needs to know about fit shows up in those few days, not in a loop of hypotheticals.

Ajit, our CEO, wrote up why we replaced interviews with this, including the uncomfortable math behind it.

SageOx Today

$15M seed led by Canaan, with A.Capital, Pioneer Square Labs, and Founders' Co-op. Founded by Ajit Banerjee (AWS EC2/EBS, XetHub → Hugging Face), Ryan Snodgrass (early Amazon, Kindle/AWS-scale cloud systems), and Milkana Brace (EVP Product at Remitly) — people who've built this kind of infrastructure, and scaled through an acquisition or two, before. We're small on purpose, hiring our first ten. More about the team →

Compensation

Competitive salary and meaningful equity. We'll talk specifics once we know it's a real mutual fit. That's what the trial is for. Staff/Principal candidates may get basic relocation help.

Apply

There's no shortage of comfortable places to write software right now. This is not one of them, and that's the point.

We're not working with recruiters or recruiting agencies/consultants at this time — direct applications only, please.

Learn more about this Employer on their Career Site

Product Engineer, Hardware