Blog
Notes from building AI projects
What I learned while building machine learning, computer vision, generative AI and data science projects, with the real numbers.
Designing an AI agent that asks before it acts
Most agent demos let the model pick any tool and run it. In AgentOS AI I went the other way: a closed tool registry, a plan you approve before anything runs, and web content treated as untrusted data.
Machine Learning · Data ScienceWhy accuracy lied about my churn model
On imbalanced data, a model that predicts “nobody leaves” scores 73.5% accuracy and catches zero churners. Here is how I evaluated my customer churn model instead: PR-AUC, recall and a business-driven threshold.
Computer Vision · Deep LearningI benchmarked 7 deep learning models on 54,299 plant images. The most accurate one wasn’t the best choice.
ResNet-50, ViT-B/16, EfficientNet-B0, DenseNet-121, VGG-16, AlexNet and LeNet-5 on the PlantVillage dataset: accuracy, latency, size, and why fine-tuning beat a frozen backbone by 8 points.
Generative AI · Developer ToolsBuilding an AI assistant that can’t make up citations
How DevMind AI combines static analysis, a dependency graph and citation checking so its AI assistant only cites code that actually exists.
Machine Learning · FundamentalsRegression from scratch: what building Ridge, Lasso and gradient descent taught me
I implemented linear, polynomial, Ridge and Lasso regression, gradient descent, K-fold cross-validation and VIF without scikit-learn, to predict prices for 10,000 real estate properties.