All

Nov 1, 2026 • 36 min read

LeNet-5 From Scratch: 61,706 Parameters Between a Photo and a Digit

A complete walkthrough of the convolutional network that started modern computer vision, built and trained from scratch in PyTorch, exported to ONNX, and run live in a browser. What every layer does, where the weights actually live, what the modern changes are worth when you measure them across eleven models, and why the export needs checking.

Machine Learning CNN Computer Vision PyTorch ONNX MNIST Deep Learning
Read Article
Sep 12, 2026 • 26 min read

Recursive Self-Improvement Is Bounded by What Can Be Checked

The case for recursive self-improvement has a hidden premise: that a machine can tell which version of itself is actually better. Sixty years of research keeps running into that one step. Where self-improving systems reached superhuman results, where they collapse instead, and what decides which, with a complete runnable Python experiment.

AI Agents Recursive Self-Improvement Verification AI Safety Reasoning Python
Read Article
Jun 20, 2026 • 48 min read

Defending AI Agents Against Prompt Injection: A Production Playbook

A deeply researched, production-grade guide to prompt injection in 2026. Why detection cannot be trusted, the lethal trifecta, and the by-design defenses that hold even when the model is fooled, with complete Python implementations of capabilities, taint tracking, plan-then-execute, the dual-LLM quarantine, egress control, MCP hardening, and a red-team harness.

Prompt Injection AI Agent Security LLM Security Agentic AI Capabilities Dual LLM Production
Read Article