Blog
I write about how I choose data, build pipelines and check results. My current AI work is a RAG learning assistant; you can explore its retrieval and evaluation workflow in the project on GitHub.
Subscribe by RSS to get new posts.
- 20 September 2026Status: In RefinementApplication: Building and publishingEngineering: Model validation
Evaluating ML, deep learning, LLMs and reinforcement learning
How to choose the evaluation unit, data split and metrics for predictions, generated answers and sequences of actions, with worked examples and a RAG case study.
- 20 September 2026Status: In RefinementApplication: Building and publishingEngineering: Data pipelines
How RAG turns documents into answers, and what to evaluate
Follow a document question through retrieval, evidence selection and answer checking, with practical choices about search, context, citations and abstention.
- 20 September 2026Status: In RefinementApplication: Building and publishingEngineering: Model validation
RAG guardrails in Python: follow a request through the code
A code walkthrough of the AI course assistant, with embedded GitHub gists for classifier routing, retrieval, OpenAI-compatible generation, grounding checks and evaluation.
- 15 September 2026Status: PublishedApplication: Building and publishingEngineering: Data pipelines
Putting NRW Charging Atlas online with Cloudflare Pages and GitHub Actions
How I turned a local geospatial application into a public map, with temporary PostGIS ETL, Terraform and a static deployment on Cloudflare Pages using GitHub Actions.
- 15 September 2026Status: In RefinementApplication: Building and publishingEngineering: Model validation
Building RAG guardrails for answers you can check
A layered RAG application that decides when to answer, block, redirect or abstain, with OpenAI-compatible model calls and a clear evaluation trail.
- 15 September 2026Status: In RefinementApplication: Building and publishingEngineering: Model validation
Try the RAG guardrails locally: a visual walkthrough
Run the credential-free pipeline, inspect a blocked request and try the local policy manager without calling a model API.
- 14 September 2026Status: PublishedApplication: Data for decisionsEngineering: Geospatial methods
NRW Charging Atlas: comparing charging coverage with GeoNode and PostGIS
How I built a district-level charging atlas with public data, inspectable scores and persistent scenarios for North Rhine-Westphalia, Germany.
- 14 September 2026Status: PublishedApplication: Data for decisionsEngineering: Geospatial methods
A quick tour of NRW Charging Atlas
Start the project locally, compare districts in the dashboard, browse the GeoNode datasets and inspect the layers in GeoServer.
- 8 April 2026Status: Published
Crude Benchmarks - API, Sulfur, and Interactive Charts
Publishing the first crude benchmark visuals for API gravity, sulfur, benchmark comparison, and regional context.
- 5 April 2026Status: Site UpdatesApplication: Building and publishingEngineering: Data pipelines
Site Inception: GitHub Pages, Astro, and a Notebook Publishing Trail
Why this portfolio exists, how it deploys, and how I am turning notebook work into public posts.
- 4 April 2026Status: Completed And AvailableApplication: Climate riskEngineering: Data pipelines
Notebook 1 - Giving the Project a Spatial Backbone
How I load the district geometry once, check it, and give the rest of the project a stable spatial frame.
- 3 April 2026Status: Completed And AvailableApplication: Climate riskEngineering: Model validation
Notebook 2.1 - Do the DWD Stations Actually Cover the Districts?
Mapping DWD climate stations to NUTS-3 districts before trusting station-based weather features.
- 2 April 2026Status: In RefinementApplication: Climate riskEngineering: Model validation
Notebook 4 - Do the Joins Hold, and Is There a Signal?
Checking join completeness and testing whether the sandy-soil drought signal is visible before the full analysis run.
- 1 April 2026Status: In RefinementApplication: Climate riskEngineering: Model validation
Notebook 5 - The Real-Data Risk Run
Feature engineering and district risk scoring with live public data, plus the caveats that keep the ranking honest.
