**Founding ML Engineer
Ideal start date is ASAP - role will be closed once a suitable candidate is
found**
**Hi! 👋**
My name is Sophia Millar. I’m the Co-Founder and CEO of Sonibel Instruments.
We’re a hardtech startup with the mission of modernizing manufacturing so we
can keep up with the world’s growing demand for urgent critical
infrastructure, from skyscrapers to naval ships to grid expansions.
**At Sonibel Instruments we work hard. You’ll get:**
* ⚙️ **Near total autonomy and freedom**. You pick your own projects, set your own goals, and work the hours you want, where you want, with who you want
* 🔧 **Only important work.** We’re a small team so there’s no busy work. You’ll be leading mission critical projects from the ground up
* 🤖 **Fast, hands-on learning** across functions. Every week comes with new problems, new skills, and real things to build
* 💡 **Decisions based on data**. Everyone at any every level is trusted to make decisions and experiment
* 💰 **Competitve compensation** through salary and **equity** so you’ll own your work
* 🦾 **Leaders who own their mistakes** , actively seek feedback to identify & solve them, and are pretty experienced. TL’DR: this isn’t my first startup; I was on the founding team of two successful startups before Sonibel. One went through Y-Combinator. Then I launched two of my own. One was profitable. My co-founders George and Hooman are both cracked engineers who invented Sonibel’s patented tech
**In the past 10 months, we’ve:**
* 💸 Raised over $1.6M USD
* 👩🏭Invented and patented a first-of-its-kind system
* 🎉Landed our first major Fortune 100 customer
* 🏆Had inbound requests from the biggest manufacturers in the world wanting to use our tool including Caterpillar, Komatsu, Seaspan, L&T, and more
* 🎓Graduated from the Creative Destruction Lab
* 😎And we just started growing the team past the 3 founders
We’re building the first ever complete real-time quality control system for
welders. Our edge device is powered by acoustics: it listens to welds in real-
time and alerts the welder if there’s a defect. This can cut total project
costs by 30% and save 2 months on a year-long timeline, making it cheaper and
faster to build urgent critical infrastructure. With a $1.5 trillion
manufacturing backlog and huge labour shortage, we’re in high demand.
**You’ll be doing and learning to:**
🧪 ML systems and experimentation
* Design and operate Sonibel's ML experiment loop end-to-end: dataset versioning, training pipelines, agent-driven hyperparameter search, evaluation, error analysis, and repeatable research infrastructure
* Run experiments autonomously by forming hypotheses, training models, interpreting results, identifying failure modes, and deciding what to try next without waiting for direction
* Maintain and evolve cloud and GPU infrastructure for training, large-scale data preprocessing, and distributed experiment orchestration
* Manage data storage, versioning, and pipeline artifacts on AWS (S3, EC2, and related services) keeping datasets, models, and experiment outputs organized, versioned, and accessible
📦 Edge deployment and optimization
* Quantize, prune, and optimize models for real-time inference on embedded hardware (ARM Cortex, embedded Linux, FPGA targets)
* Build the production pipeline: train → evaluate → package → test → deploy to edge devices in the field, with automated regression checks at each stage
* Profile latency and memory on target hardware; understand the tradeoffs between model quality and inference budget at the hardware level
🔧 Edge hardware collaboration
* Deploy and validate models on embedded hardware in the field. You own the last mile from packaged model to running inference on the device
* Collaborate with our hardware team on integration questions: understand the data path from sensor to model input well enough to debug end-to-end issues without needing to own the firmware yourself
* Flag inference requirements (memory, latency, power) early so they can inform hardware decisions upstream
🏭 Field deployment and data
* Run data acquisition sessions independently: collect high-quality welding audio, debug sensor anomalies, manage hardware in industrial environments
* Support customer pilots including first installs, sensor placement, operator training, on-site debugging, and post-deployment monitoring
* Turn field feedback, failure modes, and data quality issues into better datasets and better models; close the real-world feedback loop continuously
🔐 Data security and compliance
* Maintain awareness of data security and privacy requirements for industrial ML deployments (at-rest encryption, access controls, audit logging)
* Support compliance discussions with enterprise customers and help our CPO, Hooman satisfy security requirements during pilots and commercial rollouts
🚀 Product and strategy
* Contribute to customer-facing ML features and product decisions as one of the first people building Sonibel's core intelligence layer
* Help define evaluation standards, deployment criteria, and model governance as we scale from pilot to production across multiple industrial verticals
These will change based on what’s working best in the field. Overall, we want
someone to own the full arc from raw field data to a deployed, monitored,
improving model, working directly alongside our CTO, George and collaborating
with the hardware/firmware team when needed.
**The ideal candidate:**
* 🤖 Wants to build ML systems that work in the real world, on a shop floor, inside a noisy enclosure, on hardware with tight memory and power budgets, not just in notebooks or on benchmark datasets
* 🧪 Is comfortable running the full experiment loop independently: forming a hypothesis, writing the pipeline, training the model, reading the results, and deciding what to do next without being handed a plan
* ⚡ Cares about speed, latency, and memory as much as accuracy. A model that can't run where it needs to run isn't a model, it's a prototype
* 🛠️ Is comfortable deploying models onto edge devices and debugging what breaks between the trained model and the running device. You've poked at a sensor with an oscilloscope, or you're the kind of person who would. You don't need to write firmware, but you should understand enough of the hardware stack to collaborate effectively with the hardware/firmware team
* 📦 Thinks about the whole system: training, versioning, packaging, deployment, monitoring, field feedback. You've felt the pain of a model that worked in training and broke in the field and you know how to prevent it
* 🎧 Is excited by messy real-world data, not discouraged. Sensor noise, inconsistent environments, imperfect labels, and hardware quirks are problems you want to solve
* 🏗️ Is willing to put on a hard hat and go to a shipyard, a structural steel fab, or a construction site to understand what's actually happening. Industrial context is something you'll learn and care about
* 🚀 Takes ownership and follows through. You want real responsibility, not a backlog of small assigned tasks
* 🌟 **Bonus points** if you have:
* Experience with embedded AI (TensorRT, ONNX Runtime, NCNN, OpenVINO, or similar edge inference stacks)
* Audio or time-series ML (signal processing, STFT, mel spectrograms, acoustic event detection)
* MLOps experience (experiment tracking, model registries, drift monitoring, CI/CD for ML)
* Frontend or full-stack experience - we need customer-facing dashboards that surface model performance, defect history, and real-time alerts to operators and quality managers. If you can build that, you'd own it end-to-end and it's a core part of the product
* Industrial AI, robotics, NDT, or field deployment of ML systems
* Experience shipping a model to production and maintaining it post-launch
* Embedded Linux or firmware familiarity
* Meeting even 50–70% of these means you could be a great fit!
If you made it this far, **apply** by sending our CTO, George an email with a
few times that work for an interview:
[[email protected]](mailto:[email protected])
_This write up was inspired by our friends at Pilot!_