Machine Learning Engineer

Company Description and The role

\n\n

Philyron (a16z SR) Working with the co-founders, you will own the entire pipeline from vessel-edge connectivity (MQTT/Sparkplug B) to cloud architecture (AWS) and the ML fusion engine that predicts failure modes using sensor and historical data. Full Stack preferred.

\n

Concretely, that means everything after the hardware: getting data off the vessel, into the cloud, onto a dashboard, and through the machine learning models that make the whole thing worth paying for. You won't be wiring sensors — but you'll own everything those sensors feed into.

\n\n

What you'll build

\n\n

Connectivity — get data off the ship. Ships have terrible, intermittent connectivity. You'll build the pipeline that moves sensor data from a vessel's edge device to our cloud reliably over satellite links (VSAT / Starlink) — using lightweight, fault-tolerant protocols (MQTT / Sparkplug B), with encryption in transit and graceful handling of connections that drop for days at a time and resume.

\n\n

Cloud backbone — land it and make it usable. You'll stand up our cloud architecture (AWS): a real-time "hot path" that checks incoming sensor metrics against safety thresholds and fires instant alerts, and a "cold path" that stores high-frequency time-series data for trend analysis and model training. You'll also build the integration layer that pulls decades of historical maintenance records out of operators' existing CMMS databases — the data that makes our predictions possible.

\n\n

Predictive fusion engine — the reason we win. This is the heart of it. You'll build the ML system that fuses two data streams almost no competitor combines: live sensor signatures and historical failure records. You'll map past maintenance events to the sensor patterns that preceded them, train anomaly-detection and failure-prediction models (e.g. isolation forests / autoencoders for anomalies, gradient-boosted trees / LSTMs for remaining-useful-life), and design the fleet-wide-plus-per-vessel modeling approach that makes predictions both accurate and personalized to each engine. The output isn't "anomaly detected" — it's "this pattern preceded a gearbox bearing failure across the fleet; inspect within 14 days."

\n\n

Automated reporting. You'll build the service that compiles trends, history, and predictions into clean reports delivered to operators on a schedule — turning the platform into something that shows up in their inbox and proves its value every week.