Design long-running agents by separating durable session history, evolving harness logic, and disposable execution environments.

Anatomy of a Long-Running Agent: Session, Harness, and Sandbox

Evaluate always-loaded tools, deferred discovery, and programmatic orchestration before expanding an agent tool catalog.

Too Many Tools: Testing Agent Tool Discovery

Understand the experimental MCP Tasks extension for long-running tools, including polling, cancellation, human input, and authorization.

MCP Tasks: Durable Asynchronous Tool Calls

Measure coding-agent productivity using accepted work, review effort, and later rework instead of generated code or perceived speed.

Are Coding Agents Actually Saving Time? Measure Review and Rework

Design enforceable filesystem, credential, and network boundaries that limit what an AI agent can reach when model-level safeguards fail.

Containing AI Agents: Filesystem, Secrets, and Egress Boundaries

Build model routing and fallback policies that preserve application contracts while measuring quality, latency, and total cost.

LLM Routing and Fallback: Preserve the Contract, Measure the Tradeoff

Learn how plans, durable artifacts, and evaluator loops help coding agents complete long-running software tasks without losing direction.

Long-Running Agent Harnesses: Plans, Artifacts, and Evaluator Loops

Explore graph-based RAG systems that combine knowledge graphs with LLMs. Learn Microsoft GraphRAG, Neo4j integration, and when to use graph vs vector retrieval.

GraphRAG: When Knowledge Graphs Meet Retrieval

A comprehensive guide to productionizing AI agent solutions. Includes a complete working project example, security practices, compliance considerations (SOC2/HIPAA/GDPR), and actionable checklists for production readiness.

Production AI Agents: The Complete Checklist

OpenClaw (formerly Clawdbot/Moltbot) became the fastest-growing GitHub project ever, but its security vulnerabilities and the Moltbook social network for AI agents have created unprecedented risks. This guide covers the evolution, capabilities, and critical security concerns every developer should understand.

From Clawdbot to OpenClaw: The Rise of Personal AI Agents and Their Security Risks

Learn how to reliably extract structured data from LLMs. Covers native JSON modes, Pydantic integration, function calling, and handling edge cases.

Structured Outputs: Getting Reliable JSON from LLMs

Learn how to evaluate LLM applications effectively. Covers evaluation frameworks, metrics, test set creation, and continuous monitoring strategies.

LLM Evaluation: Testing AI Systems That Actually Work

Learn how prompt caching can reduce LLM API costs by up to 90% and improve latency. Covers implementation strategies for Anthropic, OpenAI, and custom caching solutions.

Prompt Caching: Optimizing LLM API Costs and Latency

Learn how to build production-ready AI agents using Google's Agent Development Kit (ADK). Covers agent architecture, tool integration, multi-agent systems, and deployment with Vertex AI.

Google's Agent Development Kit (ADK): Building Production Agents

Explore the Model Context Protocol (MCP), an open standard for connecting AI models to external tools and data sources. Learn how to build MCP servers and integrate them with Claude and other AI systems.

Model Context Protocol (MCP): Standardizing AI Tool Integration

Learn how to structure Python ML projects for production. Covers project layout, dependency management, configuration, testing, and CI/CD practices.

Production Python: Project Structure and Best Practices for ML

Learn how to build robust, scalable APIs for machine learning models using FastAPI. Covers async patterns, streaming, background tasks, and production deployment.

FastAPI for ML Engineers: Building Production APIs

Take your RAG systems to the next level with advanced techniques like query expansion, hybrid search, reranking, and sophisticated chunking strategies.

Advanced RAG: Beyond Basic Retrieval

A comprehensive guide to understanding, building, and deploying AI agents. Learn about agent architectures, tool use, memory systems, and production considerations.

AI Agents: From Concepts to Production

Explore multimodal AI models like LLaVA, GPT-4V, and Qwen-VL that understand both images and text, with practical code examples.

Multimodal AI: Building with Vision-Language Models

A practical guide to LLM quantization techniques for running large models on consumer hardware with minimal quality loss.

LLM Quantization: GPTQ, AWQ, GGUF and When to Use Each

Learn how to run open-source LLMs locally using Ollama and vLLM for privacy, cost savings, and low-latency inference.

Running LLMs Locally: A Complete Guide to Ollama and vLLM

A practical comparison of LangChain and LlamaIndex for building LLM-powered applications with code examples for common use cases.

LangChain and LlamaIndex: Building LLM Applications

Master the core techniques of prompt engineering—from zero-shot to chain-of-thought—to get consistent, high-quality results from any LLM.

Prompt Engineering: Getting Better Results from LLMs

A comprehensive guide to building Retrieval Augmented Generation (RAG) systems that combine the power of LLMs with your own knowledge base.

Building RAG Systems: Retrieval Augmented Generation from Scratch

Understand how vector databases work under the hood, when to use them, and how to choose between Pinecone, Weaviate, ChromaDB, and Qdrant for your application.

Vector Databases Explained: Architecture and Selection Guide

Learn how Named Entity Recognition works, when to use different approaches, and how to train custom models for domain-specific entities.

Named Entity Recognition: Extracting Structured Information from Text

A guide to text embeddings from Word2Vec to Sentence Transformers—how they work, when to use each type, and practical implementation patterns.

Understanding Text Embeddings: From Words to Meaning

Deep dive into reward modeling - the critical first step in RLHF that teaches AI systems to predict and optimize for human preferences through comparative learning and preference ranking.

Reward Modeling for RLHF: Teaching AI to Understand Human Preferences

Comprehensive guide to supervised fine-tuning of Large Language Models, covering data preparation, training implementation, hyperparameter optimization, and evaluation strategies with practical code examples.

Supervised Fine-tuning Deep Dive: Building Your First Instruction-Following Model

Complete guide to setting up a robust development environment for LLM fine-tuning, covering hardware requirements, software installation, data preparation workflows, and optimization techniques.

Setting Up Your LLM Fine-tuning Environment: Hardware, Software, and Best Practices

A comprehensive introduction to LLM fine-tuning covering key concepts, different approaches, and guidance on choosing the right method for your use case.

LLM Fine-tuning Fundamentals: Understanding When and How to Fine-tune

Understand how LoRA enables fine-tuning of billion-parameter models on consumer GPUs by learning low-rank adaptations instead of updating all weights.

LoRA and QLoRA: Efficient LLM Fine-tuning on Consumer Hardware

Comprehensive guide to OpenCV for computer vision applications, covering image processing, feature detection, object tracking, and real-time video analysis with practical examples.

OpenCV Tutorial: Computer Vision with Python

Exploring Model Soup techniques that improve neural network performance by averaging weights from multiple fine-tuned models, offering better accuracy without increased inference cost.

Model Soup: Improving Deep Learning Through Weight Averaging

Comprehensive exploration of local image descriptors including classical methods like SIFT and ORB, and modern learned approaches for keypoint detection and description.

Local Image Descriptors: From SIFT to Learned Features

Comprehensive guide to global image descriptors, exploring traditional methods like HOG and LBP alongside modern deep learning approaches for image representation and retrieval.

Global Image Descriptors: From HOG to Deep Learning Features

Understanding ConvNeXt - the systematic modernization of convolutional networks that proved CNNs could compete with Vision Transformers by adopting their best design principles.

ConvNeXt: How Classic CNNs Fought Back Against Transformers

Exploring Test-Time Augmentation techniques that improve model predictions by aggregating results from multiple augmented versions of test images.

Test-Time Augmentation (TTA): Boosting Model Performance at Inference

Understanding image augmentation - why it's crucial for computer vision success, what techniques work best, and how to apply them effectively to build robust models with limited data.

Image Augmentation: The Art of Creating More from Less

Understanding contrastive learning - the breakthrough approach that teaches AI to recognize patterns by comparing what's similar and what's different, without needing labeled data.

Contrastive Learning: Teaching AI Through Comparison

Exploring PyTorch Lightning, a framework that organizes PyTorch code for scalable, reproducible, and maintainable deep learning projects.

PyTorch Lightning: Simplifying Deep Learning Research and Production

Understanding Batch Normalization, one of the most important techniques for training deep neural networks, from theory to implementation.

Batch Normalization: Accelerating Deep Network Training

Exploring the landscape of deep learning optimizers from SGD to Adam and beyond, understanding their mechanics and when to use each one.

Deep Learning Optimizers: A Comprehensive Guide to Training Neural Networks

How ChatGPT works under the hood - from predicting the next word to engaging in human-like conversations. Understanding the magic behind large language models.

ChatGPT and Large Language Models: Understanding the Revolution in Conversational AI

Understanding Stable Diffusion - the breakthrough AI system that can create stunning images from text by learning to reverse the process of adding noise to pictures.

Stable Diffusion: How AI Learned to Paint from Pure Noise

How Vision Transformers challenged CNNs by treating images like sentences - breaking them into patches and using attention to understand spatial relationships.

Vision Transformer (ViT): Bringing Transformers to Computer Vision

A conceptual deep-dive into the Transformer architecture that revolutionized AI. Learn the intuition behind attention, why it works, and how it powers modern language models like GPT and BERT.

Understanding the Transformer Architecture: The Foundation of Modern AI

Understanding the different computer vision tasks - object detection, semantic segmentation, instance segmentation, and panoptic segmentation - their applications, and when to use each approach.

Computer Vision Tasks: From Object Detection to Panoptic Segmentation

Understanding CLIP - the breakthrough model that bridges vision and language, enabling AI to understand images through natural language without task-specific training.

CLIP: Teaching AI to Connect Images and Language

Understanding ArcFace loss - the breakthrough technique that revolutionized face recognition by teaching networks to create better feature boundaries through angular margins.

ArcFace Loss: Teaching Neural Networks to Create Perfect Boundaries

What is the purpose of WSL 2, including installation and setup for Data Science

Installing WSL on Windows 11 and setting up environment for Data Science

In this blog we go through adding images in nextjs in mdx formats

Understanding how to add react component such as Images in Next.js

Quick setup to get everything installed on your local machine for tensorflow and pytorch

How to setup Tensorflow and Pytorch for your local machine in Windows 11 with a GPU

Understanding XGBoost - the extreme gradient boosting algorithm that dominates machine learning competitions. Learn how it works, why it's so effective, and when to use it.

XGBoost: Understanding the Champion of Machine Learning

A Primer on Time Series using R and Python

Exploring historical Tennis data using R and Visualisations

ATP Men’s Tennis Analysis using the R package GGPlot2

I go through my experience attending the Melbourne Datathon 2016 and placing 4th.

Research, projects, and deployed AI solutions

Mostly thinking, building, and seeing what works.

New posts from the workshop floor.

Anatomy of a Long-Running Agent: Session, Harness, and Sandbox

Too Many Tools: Testing Agent Tool Discovery

MCP Tasks: Durable Asynchronous Tool Calls

Are Coding Agents Actually Saving Time? Measure Review and Rework