AI engineer in manufacturing

pplwise GmbH; Berlin; Hybrid; EUR 70,000 - 110,000; Full time

Job summary

pplwise GmbH in Germany seeks a machine learning engineer to design robust data splits and quotas, and to engineer automated gates for model changes. You will work on leakage-safe pipelines and per-class quotas to ensure reliable performance.

You will train open-weight vision-language models on technical documents, repair labels, and build reproducible GPU serving artefacts. Collaboration with customers and teams is essential as you scale multi-tenant deployments.

Qualifications

• You report the number that makes your model look worse, and say plainly when a result cannot answer the question.
• From dataset and training through serving, monitoring and rollback in production.
• You are comfortable with ambiguity, because defining the problem, the benchmark and the metric is the first job.
• You write decisions down with their reason and their measurement, and share progress, blockers and trade-offs early

Responsibilities

• Design leakage-safe splits, per-class quotas for the rare classes that matter, and power calculations before comparisons, then turn it into an automated gate and a durable record so every model change is adjudicable.
• Train open-weight vision-language models on technical documents and own the cheaper models in front of them, select on the task metric rather than the loss curve, and kill your own model when a simpler approach wins.
• Define label taxonomies precisely, build ground truth where none exists, repair labels with clean provenance, and close the loop so human corrections become training data.
• Own the parity between how features and preprocessing are computed at training time and in production, as executable checks rather than conventions.
• Stand up models for new customers in a multi-tenant setup, with thresholds and model identity that travel per customer and tenant isolation as a hard requirement.
• Understand which mistakes are cheap and which are expensive, review real output with customers and colleagues, and explain a model's limits honestly to the people acting on it

Skills

Ambiguity tolerance

Clear communication

Data quality review

Job description

Описание:

My co-founder Ian and I combine over 15 years of industry experience, working with companies like SEW-Eurodrive, Terex, Bosch and Carl Zeiss. We co-founded Resourcly, a Project A-backed, industrial tech start-up to help industrial manufacturers streamline complex portfolios to improve EBIT while preventing and reducing idle inventory from being scrapped.

Задачи:

Design leakage-safe splits, per-class quotas for the rare classes that matter, and power calculations before comparisons, then turn it into an automated gate and a durable record so every model change is adjudicable Train open-weight vision-language models on technical documents and own the cheaper models in front of them, select on the task metric rather than the loss curve, and kill your own model when a simpler approach wins Define label taxonomies precisely, build ground truth where none exists, repair labels with clean provenance, and close the loop so human corrections become training data Own the parity between how features and preprocessing are computed at training time and in production, as executable checks rather than conventions Take a training run to a reproducible serving artefact on GPU, measure like for like, and pick the cost lever that actually applies Stand up models for new customers in a multi-tenant setup, with thresholds and model identity that travel per customer and tenant isolation as a hard requirement Understand which mistakes are cheap and which are expensive, review real output with customers and colleagues, and explain a model's limits honestly to the people acting on it

Требования:

You report the number that makes you or your model look worse, and say plainly when a result cannot answer the question being asked From dataset and training through serving, monitoring and rollback in production Comfortable with ambiguity, because defining the problem, the benchmark and the metric is the first job You look at the labels, the data and the preprocessing before reaching for a bigger model You can inherit somebody else's code, name its strengths as readily as its faults, and improve it in place rather than rewriting it You want to know what a specification is and what the customer does with your answer You write decisions down with their reason and their measurement, and share progress, blockers and trade-offs early

Условия:

3 days in office (Mannheim, Berlin or Munich) Start: As soon as possible Compensation: Top of market (We're constrained by bandwidth, not demand - great time to negotiate)