Skip to main content

AI · LLM assistants · machine learning · data science

AI systems that turn complex business problems into practical decisions.

I design, evaluate and deliver AI products end to end - from LLM assistants and workflow agents to predictive models, risk tools and decision-support systems. Each solution starts with a clear business need and ends as something a team can understand, trust and operate.

See selected workStart a conversation

Szczecin, Poland - working remotely across the EU

What I do

AI solutions from idea to working product

I help turn business needs into useful AI tools - from knowledge assistants and workflow agents to predictive models and dependable production systems.

  • AI & LLM solution design

    I turn a business need into a clear plan for an AI product. That means identifying where generative AI can create real value, what information it needs, what could go wrong and how success will be measured - before committing to a costly build.

    • Use-case discovery & feasibility
    • Model, data and architecture choices
    • Prototype, roadmap and build / buy advice

  • AI assistants, RAG & agents

    I build AI assistants that work with trusted company knowledge, use business tools and complete clearly defined tasks. Responses are grounded in approved sources, checked against rules and designed to fail safely instead of inventing an answer.

    • Knowledge assistants and RAG search
    • Tool-calling agents and workflow automation
    • Structured answers, guardrails and evaluation

  • Machine learning & decision support

    When the goal is to predict, classify, rank risk or recognise patterns, I develop and compare machine-learning models against a strong, simple baseline. The result can be a forecast, recommendation, interactive map or analysis that supports a real decision.

    • Prediction, classification and risk ranking
    • Text, image and geospatial modelling
    • Clear evaluation and explainable results

  • Production AI systems

    A promising demo becomes useful only when it connects reliably to real data, applications and people. I turn prototypes into maintainable services with testing, monitoring, cost controls, fallback behaviour and documentation for the team that will operate them.

    • APIs, business tools and data integrations
    • Quality, latency and cost monitoring
    • Cloud deployment and handover documentation

Selected work

What I built, why it mattered

A selection of live products, research projects and data studies - explained through the problem each one addressed and the practical result.

Travel platform · AI assistant

MadeiraWise - a practical travel planner for Madeira

A map-based app that helps visitors explore attractions, walking trails, events and local weather, then turn those options into a realistic day-by-day plan. It also helps local activity providers manage bookings and guest requests.

  • React
  • Supabase
  • PostgreSQL / PostGIS
  • OpenAI
  • FastAPI
  • LangChain
Visit MadeiraWise(opens in a new tab)

Outcome

The AI planner works with real places, trails, events and weather records stored by the platform. That keeps its suggestions useful and grounded in information the product can check.

Road safety · risk mapping

Road accident risk map for Szczecin

A tool that estimates which parts of Szczecin are more likely to see a road accident during a chosen six-hour period. It combines public accident records from 2020-2025 with historical weather and patterns from nearby places and times.

  • XGBoost
  • Geospatial data
  • Weather data
  • Risk modelling
Explore the map & notebooks(opens in a new tab)

Outcome

The result is a city-wide risk map covering about 500 areas. Instead of pretending to predict a specific crash, it highlights where attention and preventive action may be most useful.

years of public accident data
2020-2025
areas shown in each city risk map
~500
time window covered by each forecast
6 hours

Known limits: This is a decision-support tool, not a prediction that a crash will happen at an exact place or time.

Wildfire planning · forecasting

How soon could a wildfire threaten an evacuation zone?

A competition project that estimated how long it could take a wildfire to reach areas near buildings. The model produced separate risk estimates for the next 12, 24, 48 and 72 hours to support staged evacuation planning.

  • Survival analysis
  • Risk forecasting
  • CatBoost
  • Model ensemble

Outcome

Testing showed that near and distant zones behave differently. I used that insight to combine four types of model, with safeguards so the predicted danger rises consistently as the time horizon grows.

forecast horizons: 12, 24, 48 and 72 hours
4
models trained and compared
~500
final combined forecasting approach
4-model
Content classification · research study

Sorting online content into advertising categories

My postgraduate thesis compared five ways to assign articles to 432 standard IAB content categories used in contextual advertising. I built a 145,415-item dataset and tested both traditional methods and a modern language model.

  • Natural language processing
  • DistilBERT
  • SVM
  • Model comparison
Read the plain-language write-up(opens in a new tab)

Outcome

The language model delivered the best overall result, assigning 87% of articles correctly and performing consistently across the 432 categories. A much lighter alternative still reached 82% accuracy, showing when lower cost and simpler operation may be the better business choice.

content categories
432
labelled text samples
145,415
approaches compared
5

Further projects

Image recognition · research study

Recognising chinchillas in photos

A study of whether an image model could reliably distinguish chinchillas from visually similar animals such as hamsters, guinea pigs and rats, as well as unrelated images.

After balancing the uneven training data and adapting a model that had already learned from a large image library, the final system reached 98.6% test accuracy and misclassified 1 of 301 test images at the chosen confidence setting.

  • Computer vision
  • EfficientNet
  • Transfer learning
  • Model evaluation
Customer experience · hackathon

What makes Shinkansen passengers satisfied?

A team project that predicted whether passengers were satisfied with their bullet-train journey by combining travel details with post-journey survey responses.

Our combined machine-learning approach placed second in the hackathon. The analysis also showed which parts of the journey were most closely linked to a positive experience.

  • XGBoost
  • Customer surveys
  • Team project
  • Hackathon - 2nd place
Image recognition · model comparison

Recognising four facial expressions

An educational project that classified face images as happy, sad, neutral or surprised. I compared four models built from scratch with three established models that had already learned from a large image library.

The best custom model matched the larger alternatives at about 71% test accuracy. The comparison supported choosing the smaller, easier-to-retrain option and clearly exposed the need for cleaner, broader training data.

  • Computer vision
  • CNN
  • Transfer learning
  • TensorFlow / Keras
Food delivery · data analysis

Turning FoodHub order data into business actions

An end-to-end analysis of a food-delivery service: what customers ordered, when they ordered, how long preparation and delivery took, which restaurants generated revenue and where ratings were missing.

The analysis led to practical recommendations for promotional offers, delivery-time targets, customer ratings and campaigns aimed at one-time customers.

  • Python
  • pandas
  • Business analysis
  • Data visualisation
Travel planning · generative AI

HealthyTravel - lower-impact, health-aware trip planning

A web app created for the Google Gemini Contest. It planned lower-carbon trips, suggested meals around dietary needs and health limitations, and let users provide trip details by voice.

The app brought itinerary planning, meal planning, personal preferences and traveller matching into one experience. It also estimated cost, calories and carbon impact so users could compare practical trade-offs.

  • Google Gemini
  • Next.js / React
  • Firebase
  • Voice input
  • Open source
View the source on GitHub(opens in a new tab)

How I work

A method, not a stack

The tooling changes every year. The sequence that keeps projects out of trouble does not.

  1. Start with the decision, not the data

    A model only matters if it changes what somebody does. I begin by writing down the decision it supports, who makes it, and how we will know the model helped - before any data is touched.

  2. Build the boring baseline first

    A rule, a heuristic, a logistic regression. Everything more sophisticated has to beat it on a metric we agreed on. Surprisingly often, the baseline is close enough to ship.

  3. Evaluate honestly, including the failures

    Correct validation splits for the data at hand - temporal, spatial, grouped. Metrics chosen for the real class balance. I report where a model is weak as clearly as where it is strong.

  4. Ship the simplest thing that works

    Smaller models are cheaper, faster, easier to explain and easier to retrain. Complexity gets added when the evaluation demands it - not because it is fashionable.

About

I turn complex technology into useful, dependable products.

I am a senior software and data engineer with more than 10 years of experience building online products and the systems behind them. My work has included e-commerce platforms, payment and booking services, and cloud systems that process millions of events each day.

I now combine that engineering background with data science and machine learning. Recent projects include a travel-planning platform, road-accident and wildfire risk tools, content classification, and image recognition. Because I have built production software for years, I think beyond the model: the result also needs to be reliable, maintainable and genuinely useful to the people using it.

I start with the real question, explain the options in plain language and choose the simplest approach that solves the problem well.

  • 10+ years across software, data and cloud engineering
  • AI and data science backed by production experience
  • Clear communication with technical and non-technical teams
  • From business need to a working, supported solution
Writing · MediumThe power of simplicityDon't let the hype surrounding AI and LLMs cloud your common sense - what a content-classification study revealed about matching the model to the problem.(opens in a new tab)

Toolbox

Modelling
  • scikit-learn
  • XGBoost / LightGBM
  • TensorFlow / Keras
  • PyTorch
  • Transformers
  • LangChain / OpenAI API
  • Statsmodels
Data & engineering
  • Python
  • SQL
  • pandas / NumPy
  • PySpark
  • Airflow
  • Docker
Geospatial & viz
  • GeoPandas
  • H3 / spatial indexing
  • Folium / Leaflet
  • Plotly
  • Matplotlib / Seaborn
Delivery
  • FastAPI
  • PHP
  • React
  • Git / CI
  • MLflow
  • Cloud (AWS / GCP)

Contact

Have a problem worth modelling?

Tell me what decision you are trying to improve and what data exists today. If ML/AI/LLM is not the right answer, I will say so - that conversation is free.

What happens next

  1. 1. A reply within two working days.
  2. 2. A 30-minute call to scope the question.
  3. 3. A short written proposal with options and effort estimates.

The decision you want to improve, the data you already have, and any deadline.

No newsletter — the message goes straight to my inbox.

Protected by reCAPTCHA; the Google Privacy Policy and Terms of Service apply.