ML Model Evaluation: Why a 99% Accurate Model Can Still Fail in Production
Published:
Blog posts
2026
Principal Component Analysis: Cutting Through High-Dimensional Data to Find What Actually Matters
Published:
1. Introduction: The Curse of Dimensionality
Data Visualization and Exploration: A Scientific Necessity
Published:
Data visualization is far more than creating aesthetic charts; it is a critical scientific tool for understanding trends and patterns. Our biological capacity to process information is physically limited, but our visual sense acts as a high-bandwidth “network cable” for the brain. While hearing has a bandwidth comparable to a hard drive and touch to a USB port, our visual senses are approximately ten times more powerful, allowing us to ingest and process complex data flows far more efficiently than through tables alone.
Feature Engineering: Teaching Machines What to See
Published:
Introduction: The Signal and the Noise
Data Preprocessing: Why Noisy and Incomplete Data Breaks Machine Learning Models
Published:
1. The Messy Reality of Raw Data
Why Knowledge Graphs Still Matter in the Age of LLMs
Published:
Large Language Models have taken the world by storm. GPT-4, Llama, Mistral — they can write code, summarise papers, answer complex questions, and even pass medical licensing exams. It is tempting to ask: if LLMs can do all of this, do we still need knowledge graphs?
Beyond Leaderboards: Why Ontology-Based Benchmarks Give a More Accurate Picture of LLM Reasoning
Published:
Everyone agrees that Large Language Models should be evaluated rigorously. Dozens of benchmarks exist — MMLU, HellaSwag, BIG-Bench, GSM8K, and many more. Leaderboards are updated weekly. New models claim state-of-the-art performance almost daily.
2022
KG Construction From Unstructured Text
Published:
Lots of valuable information is available on the web such as Twitter, legal documents, financial/sports news and scientific articles as unstructured data. Although a lot of Knowledge Graphs (KGs) including WikiData and DBPedia are made publicly available, it may be necessary to create our own KG for an analysis that we would like to perform. By converting text to KG, we can obtain new knowledge and new insights from text sources. In this blog, we will discuss what Natural Language Processing (NLP) methods and tools should be used to build KGs.