Projects from banking, commodities trading, e-commerce, AdTech and the EU public sector. Clients are described by sector.
Commodities trading · Gas & Power
Trading data platform for a global commodities trader
Challenge. Traders depended on market data from many vendors, websites and PDF reports, each with its own identifiers and failure modes.
What I did. Distributed PySpark pipelines on AWS, vendor API integrations, resilient scrapers, PDF table extraction into MongoDB, entity reconciliation by domain identifiers, and a dimensional model on Redshift with desk-level QuickSight dashboards.
Enterprise data lake, designed and owned end to end
Challenge. Large-scale e-commerce and vendor data needed one consistent analytical foundation.
What I did. Architecture of a layered data lake on S3 (raw, staged, curated), partitioning and file-format strategy, large-scale ETL with Airflow, Glue, EMR and Snowflake, data quality checks, FastAPI services and internal UIs for operational users.
Challenge. Model versions had to be compared objectively before any production decision.
What I did. An end-to-end evaluation framework built from scratch: LLM-generated synthetic test sets, relevance labelling, automated precision scoring across versions; RAG and chatbot proofs of concept, vector database selection, data architecture and governance reviews.
Python · LLMs · RAG · Qdrant
AdTech
Real-time bidding optimisation at high volume
Challenge. Keep campaign delivery within budget and volume limits under high-throughput bidding traffic.
What I did. The model supplying bid-price coefficients to the bidder through an API, and algorithmic traffic throttling to pace requests across campaigns and sources.
Python · machine learning · APIs
AdTech · Location analytics
Audience segments from location and event data
Challenge. Advertisers needed representative audience segments and brand-level consumer insights from mobile location panels.
What I did. Scraping and ingestion pipelines for external event data such as airport schedules and sports fixtures, joined with mobile location data and resampled into representative segments; stratification logic for cross-brand behaviour, for example between quick-service restaurant and coffee chain visitors.
Python · scraping · geospatial data · sampling
Banking and fintech
Fraud, lending and credit risk models in production
Challenge. Mobile and retail banking needed automated decisions they could trust.
What I did. An anti-fraud model with production scoring, ML pipelines of a lending decision engine on Azure Databricks, credit scorecards across the retail lifecycle and an award-winning in-house CRM solution.
Azure Databricks · ADLS · Spark ML · Power BI
Have a data platform to build, fix or take over?
Start with a one-hour consultation. If we go further, its price is credited against the assessment.