AI Architecture

Production ML systems from 30 years of UNIX infrastructure experience

AI architecture is not model design — it is systems engineering. This section documents production-grade machine learning architectures shaped by over 30 years of UNIX infrastructure experience, with a focus on reliability, observability, deterministic behaviour, and long-term operability in real production environments.

AI Architecture January 26, 2026 21:15

Version Control for Deterministic Systems: Git Isn't Enough

How Merkle chains, cryptographic attestation, and reproducible builds satisfy certification evidence requirements

10 min read →
AI Architecture January 24, 2026 00:30

Testing ML Systems: Beyond Unit Tests and Accuracy Metrics

A practical testing strategy for production machine learning

4 min read →
AI Architecture January 24, 2026 00:05

Cost Engineering for ML Infrastructure: What Actually Matters

Where the money goes and what to optimise first

6 min read →
AI Architecture January 23, 2026 22:00

State Management in ML Services: Beyond Stateless Inference

Architectural patterns for ML systems that need to remember

8 min read →
AI Architecture January 23, 2026 21:14

Graceful Degradation in ML Systems: When Your Model Can't Answer

Fallback strategies for production inference that fails gracefully instead of failing loudly

7 min read →
AI Architecture January 23, 2026 18:00

The Observability Blind Spot: What ML Metrics Don't Tell You

Why accuracy looks fine while your production system burns

10 min read →
AI Architecture January 19, 2026 23:00

The Certifiable-* Ecosystem: Eight Projects, One Deterministic ML Pipeline

From training data to deployed inference — bit-identical, auditable, certifiable

8 min read →
AI Architecture January 19, 2026 00:15

A Complete Deterministic ML Pipeline for Safety-Critical Systems

From training data to deployed inference — bit-identical, auditable, certifiable

10 min read →
AI Architecture January 15, 2026 22:31

WCET Analysis for Neural Network Inference

How to prove worst-case execution time for convolution, matrix multiply, and pooling operations

10 min read →

10 of 17 articles