Road 02 — The Data Scientist

Then the interesting questions became statistical.

Lead data scientist in a central manufacturing department. Computer vision on assembly lines, anomaly detection at millions of datapoints per day, generative AI that never leaves the building — all of it delivered end to end.

Standing in a plant long enough, you notice that most of the arguments are about numbers nobody has properly analysed. So I enrolled at JKU Linz for a master's in artificial intelligence and did it while working full time — mathematics, statistics, machine learning, deep learning, generative models, plus a stint tutoring the "Hands on AI" course.

Since June 2023 I have been the lead data scientist in the central manufacturing department at thyssenkrupp Automotive. The work spans computer vision on assembly lines, anomaly detection across process signals at line scale, an on-premise RAG assistant over the company wiki, hybrid physics/ML models that cut scrap, and Bayesian networks for uncertainty quantification.

What I actually do is the whole chain: acquisition, storage, labelling, model development, validation, deployment to production, and monitoring once it is live — with version control, CI/CD and MLOps around it, on the edge and in the cloud. Industrial ML dies in the handovers, so I try not to have any.

Degree
MSc AI, JKU Linz
Role
Lead data scientist
Scale
Up to 5M datapoints / day
Stack
Python · PyTorch · Azure
Capabilities

What the data half covers

01

Machine learning & deep learning

Models built to be evaluated honestly. Grouped, time-ordered validation and metrics designed around the production question.

  • Deep learning
  • Computer vision
  • Anomaly detection
  • Time series
  • Statistics
  • Optimisation
02

Generative AI

Applied LLM work under real constraints — on-premise hosting, confidential documents, and answers that have to cite a source.

  • LLMs
  • RAG pipelines
  • Local inference
  • Model benchmarking
  • Evaluation harnesses
03

Industrial AI & MLOps

The part most portfolios skip: getting a model to survive on a shop floor, for years, without a data scientist babysitting it.

  • End-to-end delivery
  • CI/CD & MLOps
  • Edge & cloud
  • APIs & microservices
  • Azure
  • Monitoring
How I work

From sensor to decision

Every project I ship follows the same five steps. Click through them — this is the honest version, including the parts that usually go wrong.

pipeline.log 01 / 05
source
sensors, cameras and machine controllers on the line
volume
up to 5 M datapoints per day
issues
clock drift, missing cycles, sensor dropouts
action
resync on the machine cycle, flag gaps, keep the raw copy

Nothing works downstream if the acquisition is wrong. Most projects die here, not in the model.

Data science work

Projects on this road

Vision, anomaly detection, generative AI and probabilistic modelling — every one of them deployed into a real production environment rather than left in a notebook.

01 2023 — 2026 · Manufacturing quality

Semi-supervised vision on the line

Computer Vision for Assembly Line Quality Control

A camera system on the assembly line that detects foreign objects and defects in real time, trained semi-supervised because a labelled defect dataset did not — and could not — exist.

Computer VisionDeep LearningAnomaly Detection
Approach
Semi-supervised
Detection
Real time
Target
Foreign objects
Defect labels
Not required
Read the case study
02 2023 — 2026 · Process monitoring

5 million datapoints a day

Anomaly Detection at Line Scale

An end-to-end anomaly detection pipeline for manufacturing process data, with an optimisation framework that evaluated 48 000 model variants against performance metrics designed for the actual production question.

Anomaly DetectionTime SeriesMLOps
Throughput
up to 5M points/day
Variants evaluated
48 000
Latency
Low, near real time
Metrics
Custom designed
Read the case study
03 2024 — 2025 · Precision assembly

Physics and statistics, together

Hybrid Physics / Data-Driven Assembly Model

A model that keeps the physical description of the assembly and learns what it cannot capture, reducing scrap and rework rates in assembly operations by 10 %.

RegressionHybrid ModellingProcess Optimisation
Scrap & rework
−10 %
Formulation
Physics + ML
Model search
AutoGluon
Fallback
Physical model
Read the case study
04 2024 — 2025 · Knowledge & support

Generative AI that stays in the building

On-Premise RAG Assistant for the Company Wiki

A locally hosted retrieval-augmented chatbot over the internal wiki, plus a benchmarking framework to compare LLMs on company-relevant prompts and data rather than on public leaderboards.

LLMsRAGGenerative AI
Hosting
Fully on-premise
Source
Company wiki
Benchmark
Company prompts
Data leaving site
None
Read the case study
05 2024 · Process understanding

Modelling what you do not know

Bayesian Network for Uncertainty Quantification

A probabilistic graphical model mapping the dependencies between process parameters, with Monte Carlo simulation used for rigorous uncertainty quantification and robust system optimisation.

Probabilistic ModellingUncertainty QuantificationOptimisation
Model
Bayesian network
Sampling
Monte Carlo
Output
Full distributions
Use
Robust optimisation
Read the case study
06 2023 — 2024 · Design optimisation

Optimisation in the loop with the solver

Genetic Algorithms Coupled to FEM

A global optimisation framework using genetic algorithms, coupling Python directly to Abaqus to automate synthetic data generation and accelerate design convergence.

OptimisationSimulationMechanical Design
Solver
Abaqus FEM
Search
Genetic algorithm
Data
Synthetic, automated
Goal
Faster convergence
Read the case study
Play with an interactive model in the Lab →
Experience

Leading the AI work in central manufacturing

Jun 2023 — present Eschen, Liechtenstein

Data Scientist — Central Manufacturing Department

thyssenkrupp Automotive

  • Lead data scientist in the central manufacturing department: deep learning, computer vision, anomaly detection, LLMs, statistics and industrial analytics.
  • Contributed to the conceptualisation and implementation of the company digital transformation and AI strategy.
  • Delivered end-to-end solutions — acquisition, storage, labelling, model development, validation, deployment to production and monitoring.
  • Applied software engineering practice throughout: version control, CI/CD, MLOps, API development and microservice integration, on edge and in the cloud.
  • Worked hands-on with sensors, cameras and manufacturing lines as data sources.
  • Collaborated with FH Vorarlberg and OST, and supervised a master’s thesis as industrial tutor.
  • Represented thyssenkrupp at Bodenseegespräche 2024, ICML 2024 in Vienna and industry fairs.
Education

A second master's, at night

2021 Linz, Austria

MSc Artificial Intelligence

Johannes Kepler Universität Linz

  • Studied while working full time.
  • Mathematics, statistics, machine learning, deep learning and generative models.
  • Tutor for the course “Hands on AI”.
  • Average grade: Sehr gut (1).
The other road

None of this works without the engineering.

Tolerances, tool wear, fixtures and metrology. See where the domain knowledge behind these models comes from — seven years of mechanical engineering, four of them in an engine plant.

Take the engineering road
Contact

Currently looking for the next problem worth solving.

Industrial AI, manufacturing data, computer vision, generative AI — or anything where engineering and machine learning have to meet in production. If that sounds like your team, I would like to hear about it.

lucamcwood@gmail.com LinkedIn Instagram CV

Mauren, Liechtenstein