SLAM / Audio ML - AI/ML Engineer in München

Our client is building an AI assistant that runs on smart glasses to support assembly-line workers in real time — streaming and processing video, audio, and IMU sensor data live, and delivering audio feedback directly to the worker. The system is already deployed on production lines at major automotive OEMs, privacy‑first by design: everything persisted is anonymised at ingestion.

Long‑term vision: the egocentric data collected is training fuel for the next generation of robot learning — positioning the company at the heart of the race toward autonomous robots for industrial deployment across automotive, aviation, aerospace, and beyond.

The role

Clients accept the system through formal tests with hard recall and false‑positive gates — every model must provably work: which weights, trained on which data, with which config, always answerable. You'll own that machinery end‑to‑end, and alongside it, help push the system's audio and/or spatial‑reasoning capabilities forward — knowing which vehicle a worker is acting on as they move between cars, using SLAM fused with vehicle identity signals.

You’ll work as a peer to the Lead ML Engineer — they own what the system should do, you own how models get built, trained, and reproduced — designing together, in the open, with a direct line to the CTO.

What you do

• Build the machinery that makes models provable — training pipelines, experiment tracking, model registry: full lineage from dataset to deployed weights, plus the evaluation harnesses client acceptance is staked on
• Solve data scarcity — simulation‑based synthetic data pipelines for anomaly classes real factories are too good to produce often
• Push the system beyond its current computer‑vision strength — deepen audio ML and/or SLAM‑based worker–vehicle association, depending on your background
• Ship at the edge — own the anonymization models the privacy guarantees depend on; optimize everything for constrained GPU/CPU budgets on factory hardware

What you get

• The ML production culture of a company, shaped by you from the start — registry, tracking, evals, your way
• Multimodal problems (vision, audio, spatial) most teams only get one of, on data nobody else has
• Your models on real assembly lines at major OEMs within weeks, with measurable stakes

Where you'll be in 12 months

Every model that passes client acceptance is reproducible from the registry. A rare‑anomaly class hit its recall gate on synthetic data. Depending on where you focus: audio is live in a deployment, or SLAM‑based worker–vehicle association is validated on a real line — likely both, over time.

Who you are

• You report the real number, especially when it's bad — client acceptance tests leave no room for flattering evals
• You build machines that build models: reproducibility over heroics
• You prefer solving a problem once, generally, over solving it five times quickly
• You explore broadly, then converge and commit
• You're creative about data scarcity — synthesis, augmentation, simulation

Your experience

Must have:

• Strong PyTorch and production ML experience (detection / classification / tracking)
• Proven industry experience in either audio ML or SLAM/spatial perception in production — you don't need both, but at least one at real depth
• Model optimization for edge hardware (e.g. ONNX, TensorRT)

Ways to stand out:

• Depth in the other of audio ML / SLAM, beyond your primary strength
• Manufacturing, robotics, or other physical‑world domains
• PhD (preferred, not required — strong industry track record matters more)