Lead Data Scientist · Machine Learning Engineer · Applied Mathematician · Researcher
Statistical Learning · ML Systems · Causal Inference · Time Series · Scientific Computing

20 years connecting mathematics, data science and software engineering to reliable decisions, working tools and measurable operational value

My work combines four complementary roles: I formulate problems mathematically, investigate them statistically, implement methods as usable software, and lead the delivery of analytical systems that support decisions. These are not separate labels for the same work: each has its own evidence in the projects and professional outcomes below.

As a data scientist, I design analyses and predictive models around explicit decisions and uncertainty. As a machine learning engineer, I translate models into tested pipelines, APIs and monitoring tools that other systems can use. As an applied mathematician, I formulate mechanisms, constraints and numerical approximations. As a researcher, I assess assumptions, reproduce experiments and document what results do—and do not—establish.

I do not default to machine learning when a simpler statistical or mathematical approach is adequate. I start with a baseline, evaluate complexity against the intended decision, and engineer the resulting solution to remain inspectable and maintainable.

The best starting point is the work itself. These examples distinguish public, inspectable implementations from reported outcomes in professional settings.

For a deeper technical review, follow the three system architecture maps and source-code evidence, twelve curated projects, or sixteen case studies. The full catalogue is organised separately so breadth is not confused with depth.

Engineering is the work between a fitted model and a dependable decision. My public projects illustrate different parts of this lifecycle—not a claim that every repository is a deployed commercial service.

In professional work, I have also led delivery of automated analytics and sensor-based decision systems. Reported operational improvements appear below; private implementation details are not presented as public repository evidence.

My work starts with problems that have operational or economic consequences: resources being allocated inefficiently, forecasts that do not support decisions, unreliable information, avoidable process costs, and risks that are poorly understood. Mathematics provides a way to express these problems precisely. Software turns those formulations into tools that can be used, evaluated and improved.

The value of a model is therefore not the algorithm itself, but the decision or process it enables:

• Costs, waste and scarce resources. Forecasting and constrained optimisation connect uncertain demand to staffing, inventory, scheduling and resource-allocation decisions. The economic question is not simply whether a forecast is accurate, but whether using it leads to a better decision under real constraints.
• Commercial and operational risk. Probabilistic modelling, calibration and explicit loss functions make the consequences of different errors visible. Decision thresholds should reflect the costs of false alarms, missed events and unnecessary intervention.
• Productivity and analytical infrastructure. Data integration, validation and automation reduce manual reconciliation and make information more consistent and usable. The purpose is to support more reliable decisions and free effort for higher-value work.
• Measurement of economic effects. Statistics, econometrics and causal inference help distinguish correlations from effects that could justify a business or policy decision. Assumptions, counterfactuals and uncertainty matter because acting on the wrong conclusion has a cost.
• Reliable operational monitoring. Statistical change detection and evidence-quality rules help distinguish genuine events from noise and unreliable observations, avoiding decisions driven by misleading signals.

In an industrial manufacturing setting, I worked across production management, demand planning, materials, warehousing, logistics and customer-service operations. The objective was to turn operational uncertainty into practical planning tools, not simply to improve a model score.

• Demand forecasting and production planning. Developed statistical and predictive models for demand across 60 products and 80 international customers. Forecasts informed capacity planning, production scheduling, material requirements and responses to customer orders, connecting time-series modelling with operational decisions.
• Inventory and logistics analytics. Built tools to study stock coverage, delivery lead times, forecast-versus-actual demand and logistics costs, supporting purchasing, replenishment, warehousing and dispatch decisions. Moving reporting and data management from spreadsheets into SQL made operational information more consistent and reusable.
• Manufacturing process and quality. Used deterministic and stochastic mathematical models of heat transfer, including partial differential equations, to study thermal behaviour during cork disinfection. The contribution was to understand process variability and its implications for production conditions and quality assurance, rather than treating process settings as arbitrary constants.

Reported outcomes from that professional work include a 12% improvement in forecast accuracy, a 40% reduction in inventory days, approximately €500K less inventory value, 25% less obsolete inventory and an 18% reduction in logistics costs. These are historical industry results; the underlying commercial data and measurement baselines are not published in this repository.

The work illustrates the connection between mathematical modelling, tools used by operational teams, and business decisions with economic consequences.

In each case, the work follows a common sequence: define the economic or operational objective, formulate the mathematical problem, build the appropriate tool, and evaluate the decisions it supports. A technically sophisticated implementation is not automatically valuable; its contribution must be judged in the context of the problem it addresses.

My earlier work in behavioural sensing and wireless-signal inference is one application of this approach, not its defining industry. Some commercial systems and measurements remain proprietary. The public projects below demonstrate related methods and software without implying that they reproduce employer systems or their reported business outcomes.

The mathematical formulation connects a real problem with a working tool. These projects show what was built and what it makes possible—not independently measured commercial outcomes.

• Problem: Customer feedback is difficult to search and interpret reliably.
• Mathematics: Retrieval relevance, grounded-answer evaluation, quality and failure criteria.
• Tool and contribution: A guarded RAG service with FastAPI and observability, enabling answers to be inspected against retrieved evidence.

• Problem: Uncertain healthcare demand must inform staffing.
• Mathematics: Rolling-origin evaluation, probabilistic forecasts, conformal intervals and resource constraints.
• Tool and contribution: A forecasting and staffing-optimisation platform that connects uncertainty to operational decisions.

• Problem: Housing-value changes and local-accommodation pressure may be confounded by geography.
• Mathematics: Spatial autocorrelation, robust regression and influence sensitivity.
• Tool and contribution: An official-data analysis pipeline that exposes associations and their limits without claiming causality.

• Problem: Population shift can be mistaken for predictive-model failure.
• Mathematics: Distribution comparison, sample-size-aware thresholds and Monte Carlo calibration.
• Tool and contribution: A statistical monitoring toolkit that separates evidence of drift from claims about model performance.

• Problem: Data-quality failures need explanations rather than opaque scores.
• Mathematics: Pattern consistency, mixed-type constraints and interpretable anomaly evidence.
• Tool and contribution: A consistency-checking library that links unusual records to violated expectations.

• Problem: Survival estimators require validation against controlled ground truth.
• Mathematics: Event-history simulation, censoring and multistate processes.
• Tool and contribution: A simulation library for reproducible method comparisons with known generating mechanisms.

• Problem: Structural changes in time series require distinct detection assumptions.
• Mathematics: Bayesian, kernel, state-space, offline and online changepoint methods.
• Tool and contribution: A unified toolkit for comparing detection approaches through a shared interface.

• Problem: Scientific model approximations must respect differential-equation constraints.
• Mathematics: PDE residuals, boundary conditions and physics-informed optimisation.
• Tool and contribution: A PyTorch framework for evaluating neural approaches to forward and inverse PDE problems.

See Featured for more projects and Case Studies for detailed problem-to-outcome narratives.

These figures are generated from the same canonical portfolio data used by Statistics, so the front page and the detailed evidence page share one source of truth.

Reported outcomes from professional delivery: 80% reduction in reporting costs · 30% reduction in analytics processing time · €500K reduction in inventory value through forecasting and operational optimisation. These figures refer to industry work, not measured results from the public GitHub projects; underlying commercial baselines are not published here.

• Failure-aware behavioural sensing — separating sensor failure, missing evidence, occupancy ambiguity, and genuine behavioural change before an alert is allowed to mean anything. The public programme lives in behavioral-sensing-research.
• Forecast → decision systems — probabilistic forecasts evaluated by downstream decisions, constraints, and operating cost rather than forecasting metrics alone.
• Causal inference and econometrics — identification, counterfactual construction, synthetic control, difference-in-differences, treatment-effect estimation, and reproducible policy analysis.
• Statistical monitoring and change detection — population drift, changepoints, calibration, and the distinction between data-quality failures and real distributional change.
• Research data infrastructure — provenance, metadata validation, integrity checks, and reproducible dataset registries in data.
• Research and teaching infrastructure — reusable LaTeX/Beamer material in academic-presentations, alongside open-source tooling across GitHub and GitLab.

→ More active research threads on Research.

My work spans statistical modelling, machine learning engineering, causal inference, optimisation, scientific computing, data engineering and applied AI. Across research prototypes and operational systems, the common discipline is to:

• Define the question and decision first. Specify the estimand, constraints, assumptions and consequences of a wrong decision before choosing an algorithm.
• Earn model complexity through evidence. Start from mathematical and statistical baselines; compare alternatives using validation that respects time, grouping, censoring and other data-generating structure.
• Make uncertainty and failure explicit. Evaluate calibration, missing evidence, drift, leakage, abstention and operating thresholds, not only aggregate scores.
• Keep the result inspectable. Use tests, provenance, frozen configurations and reproducible outputs, with enough detail to examine the claim and its limitations.

The detailed reasoning, methods and supporting projects are in Methods and Case Studies.

My background combines applied mathematics, data science leadership, machine learning engineering and statistical research. I have led data-science work in health technology, including sensor-based behavioural analysis, time-series modelling and decision-support systems. Alongside industry work, I teach as an Invited Assistant Professor at the Faculty of Media Arts and Design, Technical University of Porto (FMAD–UTP).

I am particularly interested in remote senior, lead, and principal-level work where statistical reasoning, engineering quality, and decision-making under uncertainty matter. I also collaborate on research, technical mentoring, and open-source scientific software.

For collaboration or professional enquiries, a short note describing the problem, constraints, and expected outcome is the best starting point. Longer-form technical writing and research notes live on my website and Medium.