AI Software Engineering

Building AI systems that can be evaluated, understood, and operated in production.

A model is one component of an AI application. The work around it includes context, interfaces, tools, evaluation, infrastructure, and the decisions that determine how it behaves when something fails.

That is one of the two primary lines of this blog. I write about the engineering needed to make those systems observable, testable, and useful under real constraints.

Suggested starting points

Memory Is Not Context — distinguish the final prompt from the complete agent trajectory, and verifier acceptance from independently supported answers.
MLflow for AI Engineering: Prompts Are Release Artifacts — version prompts, evaluate changes, inspect traces, and make controlled releases.
Anatomy of an MLOps Pipeline, Part 1 — orchestration and reproducible pipelines; continue with deployment and production.
AI Architecture: Training and Inference — hardware and the operational constraints around inference.
Statistical Learning: Foundations — estimation, bias, and variance behind the evaluation work.

For constrained hardware and autonomous workflows, see Edge + Agentic AI. The other primary line is narrative arcs in music: how songs and albums develop ideas, and what computational methods can tell us about that structure.

For code, recorded outputs, and scope notes, visit Experiments & materials. If your team needs to investigate its own retrieval, memory, or evaluation workflow, read about the memory and evidence diagnostic.

Published writing

All writing on AI Engineering

Prompts, evaluations, architecture, MLOps, and inference.

16 Oct 2026

When the Motif Returns, Does Its Meaning Return Too? Narrative Arcs Tested with LangGraph

The Beatles, Metallica, and Maluma featuring Farina: distinguish a semantic echo, a position that persists, and an idea contested by another voice, through close readings, embeddings, and an inspectable RAG experiment.

Music Analysis Engineering embeddings music lyrics narrative RAG

12 Oct 2026

The Evolution of Songs About Fear: A Reading Through Embeddings

Forty years of songs that talk about fear, read by a language model and measured with embeddings: the threat moved from the street into one's own mind, fear shifted to first-person confession, and songs learned to work through it.

Music Analysis Engineering embeddings music lyrics narrative semantic-axes

12 Oct 2026

Why Winners Can Look More Skilled Than They Are

Selection, survival, and reinforcement change what success can tell us. Synthetic demonstrations and experimental evidence expose the limits of learning from winners.

Applied Statistics statistics luck selection-bias machine-learning evaluation

05 Oct 2026

What Can Statistics Mean by Luck?

Being surprised and being fortunate are different judgments. A statistical account of luck must examine the reference distribution, the opportunities observed, and the boundary around control.

Applied Statistics statistics luck uncertainty evaluation

11 Sep 2026

Memory Is Not Context: Token Budgets, Narrative Relations, and Agent Economics

Context optimization is not agent optimization: a critical four-album study of retrieval, verifier failures, and whole-trajectory accounting with LangGraph and MLflow.

Engineering Music Analysis langgraph mlflow memory evaluation narrative

11 Sep 2026

What Should an Agent Remember? Album Narratives with LangGraph and MLflow

An empirical study of bounded memory, distant lyrical evidence, and token use across four albums.

Engineering Music Analysis langgraph mlflow memory narrative music

08 Sep 2026

MLflow for AI Engineering: Prompts Are Release Artifacts

Prompt versioning, evaluation, GEPA, AI Gateway, and a GCP architecture for operating AI applications with MLflow 3.16.0.

Engineering LLMOps mlflow prompts evaluation tracing genai

05 Apr 2026

AI Architecture - Notions on Training and Inference

A technical breakdown of CPU, GPU, TPU, and Edge AI hardware tradeoffs for training and inference workloads — with real-world cost data and a deep dive into Raspberry Pi 5 + Hailo-10H.

Engineering LLMOps cpu gpu tpu edge-ai inference

12 Mar 2026

Statistical Learning: Foundations, Bias-Variance and the Art of Estimation

A rigorous walkthrough of ISLP Chapter 2 fundamentals — from the formal definition of f(X) to the bias-variance decomposition, Bayes classifiers, and KNN — with Python code, real datasets, and connections to epistemology and learning theory.

Statistical Learning statistical-learning bias-variance knn linear-regression islp

13 Jan 2026

Anatomy of an MLOps Pipeline - Part 1: Pipeline and Orchestration

Part 1: Philosophy, project architecture and orchestration with Hydra + MLflow. Steps for preprocessing, feature engineering, hyperparameter tuning and model registry.

LLMOps MLOps Engineering mlops machine-learning python gcp mlflow

13 Jan 2026

Anatomy of an MLOps Pipeline - Part 2: Deployment and Infrastructure

Part 2: CI/CD with GitHub Actions, W&B vs MLflow comparison, complete containerization with Docker, and production-ready API architecture with FastAPI.

LLMOps MLOps Engineering mlops ci-cd docker fastapi github-actions

13 Jan 2026

Anatomy of an MLOps Pipeline - Part 3: Production and Best Practices

Part 3: Model selection strategies, advanced testing, production patterns, data drift, model monitoring, and production readiness checklist.

LLMOps MLOps Engineering mlops testing production data-drift monitoring

08 Oct 2025

MLflow for Generative AI Systems

September 2026 update: This article is retained as an earlier overview. Its evaluation and judge snippets do not reflect the current GenAI API. For current examples and the prompt lifecycle, read MLflow for AI …

LLMOps MLOps mlflow genai llms tracing evaluation

12 May 2025