Chris Hughes

I lead AI teams at Microsoft, helping large organisations close the gap between what they want AI to do and what actually ships. That gap—and why projects stall there—is what I think about most.

I write about applied AI, software architecture, and the discipline of building things that work. On the technical side, I enjoy understanding how things actually work—I learn by building and running experiments.

Outside of work: guitar, rock and blues, and the gym.

Views are my own.

Contents

Writing

Technical Blog Posts

Planning Multi-Agent Workflows: Trading Adaptability for Predictability

Front-loading intelligence with the Planner + DAG pattern for cost-effective multi-agent execution
Jan 19, 2026

From Orchestrator to Goal-Aware: What Happens When Every Agent Thinks

Exploring distributed coordination where agents reason about goals and route themselves
Jan 18, 2026

Architecting Multi-Agent Systems: Evolving Proven Patterns to Agentic Systems

Why agentic systems benefit from the same layered architecture as any complex application
Jan 17, 2026

Beyond Vanilla MoE: Fine-Grained Experts, Shared Experts, and Modern Architectural Innovations

Exploring architectural innovations in state-of-the-art MoE models like DeepSeek V3, Qwen, and Llama 4
Dec 15, 2024

How MoE Models Actually Learn: A Guide to Auxiliary Losses and Expert Balancing

Understanding how Mixture of Experts models balance learning across experts
Dec 1, 2024

Inside OpenAI’s gpt-oss: How Production MoE Really Works

A deep dive into how OpenAI implements Mixture of Experts architecture
Nov 15, 2024

Building Scalable MCP Servers with Domain-Driven Design

Applying DDD principles to build robust Model Context Protocol servers
Jul 31, 2024

Rethinking Object Detection as Language Modelling: Lessons from Reimplementing Pix2Seq

Insights from building a modern object detection system using sequence generation
Jun 15, 2024

A Brief Overview of Cross Entropy Loss

A refresher on one of the most commonly used loss functions in machine learning
Sep 2, 2023

YOLOv7: A Deep Dive into the Current State-of-the-Art for Object Detection

Everything you need to know to use YOLOv7 in custom training scripts
Nov 25, 2022

Effortless distributed training for PyTorch models with Azure Machine Learning and PyTorch-accelerated

A comprehensive guide to help you get up and running with the AzureML CLI v2
Dec 13, 2021

Introducing PyTorch-accelerated

A simple but powerful library for training PyTorch models with minimal boilerplate
Nov 29, 2021

Training EfficientDet on custom data with PyTorch-Lightning (using an EfficientNetv2 backbone)

A clean and clear starting point for anyone wishing to experiment with EfficientDet using PyTorch-Lightning
Jun 27, 2021
No matching items

Public Projects

Architecting Multi-Agent Systems

A reference implementation exploring three multi-agent coordination patterns, built with the Microsoft Agent Framework.

Jan 19, 2026

Func-to-Script

Turn Python functions into CLI scripts with minimal boilerplate

A lightweight and convenient tool which can be used to turn a Python function into a command line script, with minimal boilerplate!

Dec 29, 2024

PyTorch Accelerated

A lightweight library to accelerate PyTorch model training

A lightweight library designed to accelerate the process of training PyTorch models by providing a minimal, but extensible training loop which is flexible enough to handle the majority of use cases.

Dec 25, 2024

Simple PPO

Educational implementation of Proximal Policy Optimization

A clean, modular implementation of the Proximal Policy Optimization (PPO) algorithm in PyTorch, written with a strong focus on readability and educational value.

Nov 20, 2024

Pix2Seq

Object detection as autoregressive sequence generation

Implementation of Pix2Seq for object detection formulated as autoregressive sequence generation, supporting transformer and Llama architectures.

Oct 16, 2024

MCP DDD

Domain-Driven Design for Model Context Protocol servers

Demonstrates Domain-Driven Design principles applied to Model Context Protocol servers.

Oct 8, 2024

YOLOv7 Training

Clean, modular YOLOv7 implementation for custom training

A clean, modular implementation of the Yolov7 model family, which uses the official pretrained weights, with utilities for training the model on custom tasks.

Jul 13, 2024
No matching items