Data, analytics & AI.

Reliable Analytics Agent

A generic AI chatbot can give a convincing answer even when the question is unclear or the analysis is wrong. My agent is built for trust:
It is governed by solid use-case and confidence gates, and controlled by multiple deterministic guardrails. It will clarify intent and ambiguity first, then ensure that every output is grounded. In case of doubt, it will refuse to answer rather than mislead.
Try a business question on the demo’s e-commerce data; the SQL is there to inspect.

Independent demo using the public Olist e-commerce dataset. Evaluation and known limitations are documented in the repository.

Python, DuckDB, Claude, Groq, Gemini, Streamlit

Crossing the self-service analytics adoption chasm

What does it take to enable self-service analytics at scale? Delivering the tools doesn’t mean adoption will follow. This talk draws on our experience at wefox: onboarding business teams with a tailored “Whole Product” approach and DataOps support, and rethinking the role of data analytics in a data-literate organisation. The FP&A example starts at 17:38.

Recorded at CDAO Germany in April 2024, as Director, Group Data Analytics at wefox. Other speaking engagements include Big Data Minds, Berlin (2024), and Data Leaders, Utrecht (2025), on data democracy and capturing data value.

Self-driving neural network

Built for teaching my kids and anyone curious about how machines learn.
Explore various learning methods (flashcards, copycat, trial & error), observe how individual neurons impact a decision and how error correction propagates backwards. Finally, see how the trained network “self-drives”.

JavaScript, neural network implemented from scratch

Digit recognition, layer by layer

Draw a 1, then add a few strokes to make a 4. Watch the network change its prediction as you draw. Now slowly add a curve to make it a 9. Its confidence shifts towards the new answer: an ML algorithm ‘thinking’ in real time. Fun to watch, and a great teaching aid!

Visualisation and inference code by Evgeni Hasin; pretrained MNIST model from the ONNX Model Zoo (MIT).

JavaScript, pretrained convolutional neural network

VCBLRY

German vocabulary practice with AI feedback on your answers. Choose a level and a goal, practise a few words, then come back to the ones you missed. A way to work on the gap between recognising a word and being able to use it.

Next.js, Supabase, generative AI

HBLOKS

A block puzzle born out of restless hands during podcasts. What started as something to play while listening now has a few dozen dedicated daily players. There are daily challenges, a leaderboard and progress that follows you across devices.

JavaScript, serverless API, Redis, product analytics

Pandemonitor.org

A volunteer COVID-19 dashboard that brought country comparisons, a risk index and a 10-day mortality projection into one place. The methodology is published alongside the charts. Kept here as an archive; the data is no longer updated.

Development and data integration: Evgeni Hasin. Mathematical modelling: Ronen Altman Kaydar. Design: Roie Shalom.

Power BI, data from Our World in Data