I'm Scott Mike Robertson, a Machine Learning Engineer who takes ML from notebook to dependable product: data pipelines, model training, evaluation, and low-latency serving, built to be measured, monitored and maintained.
Trusted by data teams at
Most machine learning projects never leave the notebook. I've spent the last eight years making sure mine do. I sit at the intersection of applied research and software engineering: I can read a paper on Monday, prototype it on Tuesday, and have a version behind a feature flag, with proper evaluation and monitoring, by the end of the sprint.
My background spans recommendation systems, NLP and LLM applications, forecasting and computer vision, at startups and at regulated enterprises. I care about clean data contracts, reproducible training, honest offline metrics that match online results, and infrastructure your team can still run after I've gone.
A full-stack machine learning skill set covering modelling, data and infrastructure, so nothing gets lost between teams.
Custom architectures, fine-tuning and transfer learning for tabular, text, image and time-series problems.
Retrieval-augmented generation, fine-tuning, structured extraction and rigorous LLM evaluation harnesses.
CI/CD for models, feature stores, model registries and autoscaling inference with tight latency budgets.
Batch and streaming pipelines, data validation and versioned datasets that make training reproducible.
Offline/online metric alignment, A/B and interleaving tests, causal inference and drift monitoring.
GPU cost control, quantisation, distillation and ONNX/TensorRT export across AWS, GCP and Azure.
A selection of recent work. Client names are anonymised; the numbers are not.
Replaced a nightly collaborative-filtering batch job with a two-tower retrieval model and a gradient-boosted re-ranker served from a streaming feature store, so recommendations react to a session within seconds.
Built a retrieval-augmented assistant over 1.8M regulatory and policy documents with hybrid search, citation grounding and an automated evaluation suite that gates every prompt or model change.
Designed a hierarchical forecasting system combining gradient boosting with a temporal fusion transformer, with probabilistic outputs feeding directly into inventory optimisation for 120 warehouses.
Trained and distilled a segmentation model for surface defects, quantised it to INT8 and deployed it on edge GPUs with an active-learning loop so line operators label only the frames the model is unsure about.
What CTOs, heads of data and product managers say after working together.
“Scott took our recommender from a research demo to a system serving millions of users in under three months, and left us with the cleanest ML codebase we own.”
“The evaluation framework he built is now how our whole company decides whether an LLM feature is ready to ship.”
“Rare combination: strong on the maths, pragmatic about trade-offs, and great at explaining both to non-technical stakeholders.”
“Cut our GPU bill by more than half without touching accuracy. The project paid for itself in the first month.”
“He mentored two of our junior engineers into confident ML practitioners while delivering on every milestone.”
Every role shipped a new version. Here's what changed.
Pick the engagement that matches where you are. Every plan starts with a free 30-minute discovery call.
An expert review of your models, data and ML infrastructure.
Fixed scope · fixed price
Request an auditA production ML system, designed, built and shipped end to end.
Milestone-based delivery
Start a build →A senior ML engineer inside your team, on an ongoing basis.
Monthly rolling contract
Talk about itCan't find what you're looking for? Send a message and I'll reply within one business day.
Contact Scott →Mainly recommendation and ranking, NLP and LLM applications (RAG, extraction, classification), forecasting, and computer vision. The common thread is that the model has to run reliably in production and move a business metric.
I'm based in Warsaw and work remotely with teams across Europe and North America, overlapping CET business hours. On-site workshops and kick-offs can be arranged when they add value.
Python first: PyTorch, scikit-learn, gradient boosting libraries and Hugging Face for modelling; Spark, Airflow and dbt for data; Docker, Kubernetes, MLflow and Triton for deployment. I adapt to your existing cloud (AWS, GCP or Azure) rather than forcing a new one.
Before any modelling we agree on one primary business KPI and the guardrail metrics around it. Every model then ships with an offline evaluation report and an online experiment plan, so success is measured, not assumed.
Absolutely, and that's often the best setup. I help data scientists get their models into production, set up shared tooling, and run pairing sessions so the team owns the system confidently after the engagement ends.
Audits can usually start within a week. Build and Embedded engagements typically begin two to four weeks after the discovery call, depending on current availability.
Tell me about your data, your goals and your timeline. I'll come back with an honest view of what ML can do for you, and what it can't.