Principal Data Scientist · Builder
Anil Kumar Panda
Seeking Zima
Writing about ML, building products with AI, and the overlap between systems thinking and creative work. Based in Weesp, Netherlands.
This page is unique to you. The art you see was generated from the exact moment you arrived, your device, and chance. No one else will ever see this exact composition.
Writing
View all →Monotonic Constraints for Boosting Models
When high performance creates non-intuitive models — enforcing domain knowledge without sacrificing accuracy.
ML Model Metric Credibility
How confident are you that your model performance isn't below 0.75 AUC? A framework for honest evaluation.
Finding Where Your Model Fails
Looking at misclassified data points and inferring why your model fails on specific subpopulations.
Lab
View all →MAGA Agent
buildingMemory-Augmented Generation Agent. Finds non-obvious connections between articles using dual embeddings: topic distance × mechanism similarity = hiddenness score.
Podlist
liveConverts articles to audio podcasts. Features MAGA Agent — cross-domain connection discovery via dual embeddings and hiddenness scoring.
FlipRadar
liveMulti-model AI deal evaluator for Mac reselling — parallel evaluation via GPT-4o, Gemini, and Claude with Lazy Seller Index and Message Composer.
Kassa
liveGrocery budget insights from Dutch supermarket receipt photos. Claude Vision API + Recharts visualization.
Notes
View all →Claude writes most of the code. The interesting work is still deciding *what* to build and *why*.
Velocity beats margin in reselling. Same principle applies to features — ship fast, learn fast.
The hardest part of a PO role isn't prioritisation. It's translating between two groups who both think they're being clear.
hiddenness = topic_distance × mechanism_similarity. Two things that look unrelated but work the same way.