Machine learning and artificial intelligence

Artificial intelligence is more than chat. Many business problems are better solved by a well-chosen machine learning model: a forecast, a classifier or a search for similar cases. I help choose the approach, build the solution and check on data whether it really works.

Machine learning on data

• preparing and validating large time-series datasets and testing on historical data (backtesting),
• time-series forecasting (LSTM networks, transformers, simple statistical models as a baseline),
• classification on tabular data (scikit-learn, tree-based models, neural networks),
• detecting patterns, similarities and groups in data (clustering, DTW, nearest-neighbour search),
• reinforcement learning for decision problems, at the simulation and experiment stage,
• matching records from different sources: rules on identifiers and, for uncertain cases, probabilistic matching and manual review.

Image, speech and video

• recognising text in images and video using cloud services, and working with the output of existing object-detection models,
• speech transcription with existing models (such as Whisper) and extracting information from recordings,
• turning video analysis results into events an application can use.

Language models (LLMs)

• agents that use your company's tools and data,
• answering questions based on company documents, built on managed services (such as Amazon Bedrock knowledge bases),
• extracting structured data from text, for example from transcripts and documents,
• open-weight models running locally (for example through Ollama) when data must not leave the company; I start with a prototype.

How I work with data

• I start with the data: cleaning, validation and feature preparation,
• I evaluate every model on data it did not see during training and compare it with a simple baseline,
• I show the result in numbers before you decide on deployment,
• I connect the model's output to your application, usually through an API on AWS,
• a deployed model needs looking after: quality monitoring and retraining are agreed as a separate scope.

Experience

I built my first neural networks at university: my engineering thesis at AGH (2014) was about dynamic neural networks in modelling metallurgical processes, and I presented the project at the 49th Student Session of AGH Scientific Circles. I came back to the topic in 2018 with my own neural network written from scratch, and since 2024 I have worked on models regularly in TensorFlow and PyTorch: from transformers and reinforcement learning to a language model built from the ground up (attention, layers, tokenisation, test training runs). This is mostly my own research, largely on financial time series.

In commercial projects I worked in teams that built systems around existing models. For Sky this meant processing the results of machine-learning video analysis and recognising text on frames from live broadcasts with a cloud service. In the CMS for hayu (NBCUniversal), content metadata is enriched by an image recognition service (Amazon Rekognition). In 2026, in a short project for another client, I extended a tool used by an LLM agent. So far I have built solutions with language models and speech transcription as prototypes.

At Merapar I prepared and ran a series of internal workshops on AI and machine learning for engineering teams: data processing, feature engineering, deploying language models and generative AI architecture.

How to start

Describe the problem and the data you have. I'll tell you whether machine learning makes sense here and propose a first, small experiment.