October 12, 2026 · RAG & LLM Systems
I tested Jev, a new model from TypeSafe AI that answers classification questions with probabilities instead of generated text, on 140 made-up visit notes from a fertility and gynecology practice. It labeled 138 correctly in 2.2 seconds for less than half a cent, and with a review threshold of 0.8 both mistakes went to a person while the other 132 notes were all correct.
October 1, 2026 · Computer Vision
A nuclei segmentation pipeline on public microscopy data, from checking the hand-drawn masks to a web app that runs on a CPU, and what the trained model found in the labels it learned from.
September 15, 2026 · Explainable AI
Class activation maps are almost always explained on images. This post applies CAM and Grad-CAM to a spatio-temporal graph network for human skeleton data, with the math, PyTorch code, and the architecture detail that decides whether the two methods give the same answer.
September 10, 2026 · RAG & LLM Systems
I planted fake web pages in the sources my research agent reads, some with a false EU AI Act date and some with instructions to AI tools. The agent ignored every instruction, but the false date reached most approved reports, usually with a warning, and three fake legal pages pushed its confidence label to high. One change to how the agent judges a website brought the label down to low.
September 9, 2026 · RAG & LLM Systems
My research agent marks every claim as high, medium or low confidence, based on its sources. I had 156 of its claims checked against official sources. The low-confidence claims were right as often as the high ones, most wrong claims cited good sources but got a detail wrong, and the model's own stated confidence did better than the source-based score.
September 8, 2026 · RAG & LLM Systems
A step-by-step walkthrough of a question-answering system over the EU AI Act: scoping it, writing the test first, measuring retrieval, checking every answer, and changing one thing at a time, with the results of each step.
September 8, 2026 · RAG & LLM Systems
An AI research agent's reports read well whether or not they are right. These are the checks I built into an agent that writes cited reports on the EU AI Act, how I tested them with planted errors, and what they still cannot tell you.
July 4, 2026 · Time-Series & Signals
Before calculating the heart rate or HRV from your time series data, you need to know whether the signal is ready for further processing or modeling. This post shows a one-metric quality check for four common physiological signals, tested on real data.
June 6, 2026 · Computer Vision
A traffic sign experiment on whether synthetic training data can improve adversarial robustness without sacrificing accuracy on real images