Atoms is the industrial-robotics company Travis Kalanick built by folding CloudKitchens and Pronto into one holding structure, and in July 2026 it raised $1.7 billion led by Andreessen Horowitz. Per the July 22, 2026 announcement as reported by TechCrunch, the round also drew Uber, Bain Capital, Fifth Wall, Chemistry, K5 Global, SV Angel, and Alpha Square Group, with Ben Horowitz joining the board. No valuation was disclosed, which matters for comp later. a16z’s jobs writeup put the hiring push at 200+ open roles, so they’re staffing up fast.
A quick disambiguation: this is the Kalanick company at atoms.co building task-specific robots for heavy industry, not the shoe brand or habit-tracking app.
Three divisions, and why they change the questions
Atoms is organized into three parts: Atoms Food (infrastructure that grew out of CloudKitchens), Atoms Mining (more productive mine sites), and Atoms Transport (described as a wheelbase for robots). Feeding into all of this is Pronto, the heavy-industry haulage-autonomy company previously run by Anthony Levandowski; the same funding coverage reported that Kalanick acquired it in March 2026 and folded it into the group. The a16z bet, in Ben Horowitz’s framing, is on specialized robots over humanoids: purpose-built machines that do one industrial job well.
For a candidate, that framing is the whole tell. They aren’t hiring people to chase a general-purpose humanoid demo; they want engineers who can make a specific machine do a specific job reliably on a mine site or a kitchen line, where dust on a lens and a network link that drops mid-task are the normal operating environment. The interview rewards scar tissue from shipping real hardware over a clean benchmark number. That puts Atoms in the same cluster as Physical Intelligence and Skild AI on robot learning, Waymo and Applied Intuition on perception and autonomy tooling, and Mach Industries on shipping industrial hardware; the AI-startup interview difficulty index shows how those loops rank.
The loop, reconstructed
To be clear about method: Atoms is months old and I found no public write-up of its interview from anyone who went through it. The table below is reconstructed from the loops comparable perception and robotics companies run, cross-checked against Atoms’ actual job postings for what each role demands. Treat it as the shape to expect, not a transcript; the Source column marks what’s anchored to a posting versus inferred.
| Atoms interview stage (reconstructed) | What it screens for | Illustrative question type (not sourced) | Source |
|---|---|---|---|
| Recruiter screen | Embodied-systems background, why heavy-industry robotics, location and work-authorization fit | What hardware you’ve shipped and how it behaved in the field | Inferred from comparable robotics loops |
| Hiring-manager call | Depth in your specialty, ownership on a real project | A project you owned end to end, including where it failed | Inferred from comparable robotics loops |
| Technical phone screen | Track-specific: perception/CV, controls, or systems coding in Python and/or C++ | A perception or systems problem on a noisy sensor feed | Confirmed skills from SF perception posting; format inferred |
| Onsite / virtual panel | Domain depth, a coding or design round, hardware-debugging judgment | Debugging a machine that regressed after a firmware or data change | Inferred from comparable robotics loops |
| Values / mission round | Bias to action under ambiguity, why Atoms over adjacent offers | Why industrial robotics, and why Atoms specifically | Inferred from comparable robotics loops |
Two things the postings don’t settle, so ask the recruiter on the first call. Visa sponsorship isn’t stated anywhere public; treat it as an open question for an SF hardware startup at this stage rather than assuming either way. Leveling is the other gap: the perception role is posted at a senior band (3-5 years) and I found no public staff or principal listing, which doesn’t mean the ladder stops there. If you’re past senior, ask directly whether a staff or principal track exists.
Perception: the track with the most confirmed detail
The Senior Perception Engineer posting in San Francisco is the clearest signal Atoms has published. It asks for 3-5 years working on perception systems and computer-vision or machine-learning models, strong software engineering in Python and/or C++, and prior experience building perception for autonomous vehicles, robotics, drones, industrial automation, or other embodied systems. That last clause is the filter: web-app engineers who treat a camera as an API struggle here, while people who’ve fought sensor noise in the field do well.
Expect the technical rounds to pull from real perception work rather than puzzle problems:
- Detect and track an object across frames from a camera that occasionally drops or blurs, and explain how your tracker recovers.
- Fuse a camera and a depth sensor (LiDAR or stereo) that disagree. Which do you trust, and when?
- Your detector nails the objects in your labeled dataset but keeps missing rock fragments on a live mine feed. What’s different about the real data, and how would you close the gap?
- Calibrate a camera mounted on a vibrating platform. What drifts, and how do you catch it before it corrupts downstream control?
- Write the tight loop in C++ and explain where a memory copy or a Python GIL stall would blow your latency budget.
The through-line is the domain gap: a mine and a kitchen are hostile to a model trained on clean data, so reason about distribution shift, calibration drift, and graceful failure rather than reciting architectures. Keep your Big-O and complexity analysis sharp too, since a slow perception loop is a broken one.
Controls and robotics engineering
Atoms lists robotics roles beyond perception, including a Robotics Engineer in Madrid. It hasn’t published a syllabus, so the topics below are what a controls or robotics round of this kind generally tests, not confirmed Atoms questions. Expect state estimation to anchor it: how a Kalman or extended Kalman filter fuses noisy sensors, what’s in the state vector, and what happens when process noise is mistuned or a measurement arrives late. Motion planning comes next, sampling-based versus optimization-based planners and when each breaks on a real machine. The tradeoff interviewers most like to poke at is model predictive control versus a tuned PID loop: MPC handles constraints and lookahead but costs compute and a model you have to trust, while PID is cheap and predictable until the plant stops being roughly linear. Underneath sits real-time behavior, so be ready on deterministic scheduling, jitter, and why a missed control deadline is a safety event on a haul truck, not a dropped frame.
The other half of this track is debugging under uncertainty. A common shape: a machine that used to complete its cycle drifts off target, and you have to work out whether it’s calibration, a sensor fault, a firmware regression, or a control policy that shifted after a data change. They’re watching how you narrow a messy, poorly-instrumented failure to one cause without flailing. If you’ve tuned a controller that oscillated on the bench and settled on the machine, or the reverse, bring that story; sim-to-real scar tissue reads louder here than a clean whiteboard solution.
The residency, if you’re early-career
The rest of this guide assumes shipped-hardware experience, which would shut out new grads if it were the only door. It isn’t: Atoms runs an early-careers residency for people with 0-3 years across perception, controls, robotics, and physical-AI systems. No public reports cover how it screens, so this is inferred: early-career robotics pipelines usually weigh coursework and research depth, hands-on projects (a robotics club, a research lab, a serious personal build), and fundamentals in linear algebra, probability, and controls over years of shipping. Aim for a project you can defend down to the math, not a resume padded with frameworks.
Platform and software
Atoms names software as one of its main disciplines alongside AI, hardware, and operations, and CloudKitchens was a software-heavy operation before the rebrand, so a platform and infrastructure track almost certainly exists. I’m labeling the specifics here as inferred, since I found no dedicated platform posting to anchor them. Expect this track to look the most like a conventional senior software interview, with the twist that the data is high-rate telemetry and the consumers are often other machines. General system design interview prep transfers well, but come with one fleet-specific tradeoff worked out rather than a list of pipeline components. Take bursty telemetry: a fleet goes quiet in a dead zone, then dumps a backlog when the link returns. Do you buffer on the device and risk losing it on a power-cycle, or push hard and let ingestion shed load? That forces a real choice on delivery semantics. For a maneuver command to an actuator, at-least-once delivery means a duplicate could re-issue a motion the machine already made, so you need exactly-once or an idempotent command protocol. That is what “stricter correctness than a web backend” concretely means: a dropped web request you retry and forget, but a reordered command to a physical actuator can damage the machine or hurt someone, so ordering and idempotency stop being optional.
What the values round probably tests
I can’t confirm Atoms’ non-technical round, but given how Kalanick companies run, expect one that probes for people who move under ambiguity and own outcomes. Have concrete stories ready, kept tight with the STAR method: something you owned end to end, a time you were wrong and caught it yourself, a time you shipped against a hard deadline. Assume the “why industrial robotics” question is sincere; candidates in this wave often hold several offers, and a generic answer about loving robots lands flat against someone who can say why mining or food automation is worth their next few years.
Comp, without inventing numbers
I won’t quote a band, because Atoms hasn’t published comp and the round closed without a disclosed valuation, which makes the equity side genuinely unvaluable from outside. Two levers you can actually pull. First, the San Francisco perception role should carry a California-mandated base-salary range on the posting itself, so read that number directly rather than trusting a third-party estimate; the Madrid role sits on a different band and currency. Second, the equity is the piece that matters most and the hardest to judge with no public mark. At an offer, ask for the option or RSU count, the total shares outstanding, the strike price, and the most recent 409A valuation, then model it as a range with a total-comp calculator. If you’re weighing Atoms against other offers, walk through a real salary negotiation plan first, since early-stage companies in this wave usually have more room on equity than on base.
One last thing before you apply: make your embodied-systems experience legible on the page. Strong hardware and perception people often bury field work under generic software bullets, so run the resume through an ATS checker and confirm the robotics and perception keywords survive parsing. At a company this young, the candidate who can defend one real machine they made work beats the one who skims all three tracks.
Practice the behavioral round:
