Full-Stack Robotics Engineer (Forward Deployed)

Synphony · Bay Area, on-site with customers, heavy travel · Food & Protein · Manufacturing · Electronics · Agriculture · Logistics

The job in one sentence: Show up at a farm, a cut floor, a cable plant, or a mine, find the job nobody has been able to automate, then design the robot that does it, build it with your own hands, and put it into production.

What we are: Synphony is the deployment layer for physical AI. Already having launched from Y Combinator with multi-million dollar deals, we take frontier models—VLAs, foundation policies, and agents—and make them work inside the messy physical industries that run the real economy. Not in a lab, not in a video. In heat, dust, vibration, bad lighting, legacy PLCs, and in front of people who have done the job by hand for thirty years and do not care about your architecture. The company that owns the connective tissue between the model and the floor wins. That's the job.

The role: You get dropped into a customer with an expensive manual task and own the whole arc. Scope it on their floor with their engineers. Design the machine and build it with your hands in our warehouse. Instrument the cell, collect the data, train and fine-tune the policy, fight sim-to-real, wire it into their systems, and stand next to it on the line while it runs a shift. Then find out what you got wrong, and fix it. No clean spec, no perception team, no controls team, no one to hand the failure to. You and the task.

What you'll actually build.

The machine. Mechanism, kinematics, frame, end effector, tooling, fixturing. Machined, printed, and wired in our warehouse, by you. You're not bolting a catalog arm to a table. You decide what the machine is, and it has to survive washdown, caustic, dust, heat, and vibration, plus whatever their regulator requires.

The intelligence. Teleop rigs, camera calibration, synchronized capture, dataset curation. Fine-tuning and deploying vision-language-action models (π₀/π₀.₅-class, OpenVLA, whatever is state of the art next quarter) against real task data. Imitation learning from teleop, RL and residual policies where imitation stalls. MuJoCo and Isaac scenes, domain randomization, and honest offline eval so we know a policy regressed before the customer tells us.

The integration. Real-time control loops, safety interlocks, PLCs, line control, MES, ERP, and the operator dashboard the business actually runs on. No API, no docs, and an incumbent vendor with no reason to help you. Most of the contract value lives here.

The handover. Commission it on a live line, train their operators, instrument it so you can debug it from 1,500 miles away, and stay on the hook for uptime after you fly home. Then do it again, next customer, different industry.

You're also expected to track the frontier and reproduce what matters. We fund the compute and the hardware; you bring the judgment about what's worth trying

The bar. You've probably done several of these:

• Designed and built a working machine, not a simulation, and watched it run somewhere you didn't control it. Machined, printed, or wired your own parts because waiting on someone else was slower.
• Trained and deployed a real learned policy on real hardware, not just in sim, and got it to a rate the customer would accept.
• Hands-on with imitation learning, RL, or VLA fine-tuning: reward shaping, offline RL, DAgger, LoRA, action tokenization, inference latency budgets on edge hardware. Honest about where each one falls over.
• Worked contact-rich: force control, compliance, calibration, deformable or irregular objects.
• Wired something into equipment you didn't own: PLCs, line control, MES, ERP. No API, no docs, no help.
• Stood up everything around the machine yourself: ROS or a vendor SDK, cloud, GPU boxes, data storage, and the dashboard the customer logs into.
• Owned a machine through a real acceptance test, hit a cycle time somebody was paying for, and kept it running after handover.
• Sat across from someone who's run that line for thirty years and turned "this costs us $4M a year in labor and turnover" into a working machine.

You don't need all of it. You don't need our industries, a PhD, or a résumé that lines up. Some of the people we want haven't built any of this yet and simply figure things out faster than anyone around them.

What isn't optional: 

Raw building ability, autonomous creative problem solving, professional fluency in English because you'll be explaining yourself on a loud floor to people who don't have to listen to you, and never saying "that's not my area.”

The honest part:

You'll travel. You'll be on-site somewhere hot, loud, dusty, and far from good coffee. Robots break in ways sim never showed you. A policy that hit 95% in eval will hit 60% on their line because their conveyor runs 15% faster and nobody wrote that down. Data will be a disaster. People will be skeptical until you earn it. If that sounds miserable, this isn't your job. If it sounds like the most fun you could have as an engineer, let’s talk.

Why it's worth it:

Most robotics engineers spend a career on one subsystem of one machine in one industry. You’ll encounter more surface area and ownership than any normal role would ever hand you. At the customer site, you are the company. You'll work the entire physical-AI stack, from raw sensor bytes to a signed deployment. Every cell you stand up feeds real-world operational data back into models that get better across every other deployment. You're not shipping one robot. You're compounding a moat in the industries that grow the food, dig the materials, and make the things—the dirty jobs nobody else will touch. That's the point.