Recent

AI Systems When a Graph Earns Its Place in RAG GraphRAG builds an entity graph over your documents and retrieves subgraphs instead of passages. It wins on multi-hop and whole-corpus questions, and loses on everything else, at two to three times the query latency and a large one-off indexing cost.
Security & Resilience Are You a Data Controller or a Processor? Two regimes govern analytics data and they test different things. PECR asks whether you stored anything on the device. GDPR asks whether you can identify anyone. A 2025 CJEU judgment changed the answer to the second, and the first is still where most systems fail.
AI Systems Instrumentation Is the Deliverable The records written to debug a generation pipeline describe the process the pipeline replaced. That description is the artefact most organisations lack, and it can be produced before any model is bought.
Developer Tools Marigold: A Self-Hosted Inference API for Open-Weight Models An OpenAI-compatible API for open-weight models, running entirely on your own hardware. Open source, installed with pip, airgapped once your models are cached.

Technical Articles

Structured knowledge from client work, explorations, and observation.

AI Systems (29)

When a Graph Earns Its Place in RAG GraphRAG builds an entity graph over your documents and retrieves subgraphs instead of passages. It wins on multi-hop and whole-corpus questions, and loses on everything else, at two to three times the query latency and a large one-off indexing cost.
Instrumentation Is the Deliverable The records written to debug a generation pipeline describe the process the pipeline replaced. That description is the artefact most organisations lack, and it can be produced before any model is bought.
Data Methods for RAG Chunking is one point on a spectrum of methods for preparing data for retrieval, not the whole spectrum. A working enumeration, from fixed-size splitting through knowledge graphs to retrieving directly from document images.
RAG Risks and Costs A RAG pipeline sends your documents to a third party twice on every query, not once, and its bill scales with corpus size in a way most budgets never account for. A third risk, permission-blind retrieval, rarely gets named at all.
What Problem Does RAG Actually Solve? RAG is one answer to a specific problem: a model that does not know your data. Fine-tuning and plain search are the other two. A framework for choosing between them before any vector database gets picked.
Agent Risks, Costs, and Governance RAG's main risk is data leaving the building. An agent's main risk is different in kind: it can take the wrong action, not just retrieve the wrong document, and its cost scales with steps rather than tokens.
Evaluating RAG RAG evaluation is not one number. It is four separate measurements, taken in pipeline order, because a fault at any stage produces the same symptom: a fluent, wrong answer.
How do Neural Networks Learn? Backpropagation Explained F learns the mapping from X to Y by adjusting its weights against the error between what it predicted and what the data says was correct. Backpropagation is the mechanism that turns that error into a weight update, one layer at a time.
What Is Machine Learning? F(X) = Y, Explained Machine learning reduces to one expression: F(X) = Y. What that means, and how the four types of machine learning differ in which part of the equation is known before training starts.
Open Datasets for Training Vision-Language Models The datasets behind open vision-language models, from 108,000 hand-annotated images to 12.8 billion web-scraped pairs. Observation counts, disk sizes, and what's actually downloadable versus URL-only.
Model Choice Determines VLM Inference Cost Vision-language models are biased toward a characteristic output length regardless of image content. Because cost tracks output length, model choice determines inference cost. Benchmark data from four open-weight VLMs across 4,004 requests, with per-request power and GPU telemetry.
Why Recommendation Algorithms Resurface Old Content Recommendation systems deliberately surface underexplored content to find new audience clusters. Understanding why this happens changes how you think about content distribution.
What Is NER and Why Does It Matter for PII? Named entity recognition is the foundational step in any pipeline that handles sensitive text. What it does, where it fails, how to evaluate detection and redaction accuracy, and why it does not solve the privacy problem on its own.
Memory in Agentic AI The types of memory available to an AI agent, how they work, when each is appropriate, and what breaks when the wrong type is used or the memory layer is absent. Includes a survey of named systems: A-MEM, Karpathy's LLM Wiki, Letta, Zep, Mem0, and Cognee.
Context is a Build Step Every approach to context generation -- from basic chunking to knowledge compilation -- is an instance of the same pattern. The quality of the build step determines what your agent can do at runtime.
How to Evaluate an AI Pipeline Practical methods for scoring classification, RAG, and generation pipelines -- with dataset references by task type and domain.
Embedding-Based Relationship Discovery: Finding the Real Network in Your Data Most organisations have a formal model of their customer relationships and a real one that differs from it. Embeddings and community detection surface the real structure. Here is how the pattern works and where it applies.
Format Choice and LLM Performance: A Taxonomy The format you use to pass data to a language model affects reliability and cost more than most practitioners expect. This is a taxonomy of the main options and the conditions under which each performs well.
Task Specification is the Primary Artefact Foundation models changed the cost of the mechanism. They did not change the first question: where are the labels? A labelled dataset is a measurement instrument, not training fuel.
Static vs Dynamic Planning in AI Pipelines Most production AI workloads are high-volume and fixed-task. Dynamic planning adds cost and reduces auditability without adding anything a static pipeline cannot do.
Agent Alloys or Persona Leakage? Context bleeds between agents in multi-agent systems. The XBOW experiment suggests that for exploratory tasks, this is a feature rather than a failure mode.
Why Model Architecture Matters Every neural network architecture encodes assumptions about the structure of the problem. Those assumptions determine what the model can learn -- not just what it learns from.
Filtered Vector Search Vector search in production needs metadata filters, business rules, and combined ranking signals. How index structure determines what actually executes efficiently.
LLM Training Fundamentals: From Tokens to Human Preference LLMs are trained in stages with different objectives. How the shift from statistical training to human preference ranking produces behaviours that metrics alone cannot explain.
Sampling Strategies for Imbalanced Data Imbalanced datasets cause models to ignore the minority class. The sampling strategies that actually work for fraud detection, medical diagnosis, and quality control.
LLM Evaluation Metrics LLM outputs are probabilistic and context-dependent. A structured approach to evaluating language models in production across multiple dimensions.
A Complete Guide to Categorical Encoding for Machine Learning One-hot, ordinal, target, embedding -- categorical encoding choices affect model performance significantly. A practical guide with model-specific recommendations.
Why Are Vector Databases Difficult? A Deep Dive Vector databases are harder than they look. A technical examination of similarity search, indexing trade-offs, and why naive implementations fail at scale.
The Components of RAG RAG is not one architecture but four independent decisions: how you chunk, how you index, how you retrieve, and how you generate. A component breakdown, with the tooling that matches each one.

Product & Strategy (20)

Engineering in the Exponential Age: MCP as a Case Study MCP moved from a genuine differentiator to an assumed interface faster than REST did. A worked example of what that shift actually requires an engineering team to build, and what it returns.
How to Ensure Action Conversation between an entrepreneur and a technical person is free and valuable. The moment something gets built, it needs a company and a contract behind it, or the relationship drifts somewhere nobody chose.
How to Structure an ML Project ML projects open with a sequential setup phase before iteration can begin. What that phase involves, why it does not compress, and how the four project phases fit together.
Engineering in the Exponential Age Capability is arriving faster than teams can absorb it. The engineering task becomes assembly, not construction: protocols and interfaces in place of bespoke services.
When Training is a Commodity Inference is already a commodity. Training is next. When it arrives, every organisation that has spent the intervening period curating its own data will have a model. The AI provider will be as unremarkable as the web host.
A Measurement Framework for Machine Learning Projects A practical framework for defining, building, and maintaining the measurement system that determines whether an ML project is working. Includes scoring formulae, risk weighting, and a readiness checklist.
Don't Build Models, Build Measurements Most AI projects fail before the model is built. The question that determines whether a project is viable is not which model to use, but whether the organisation can define what correct looks like.
What Software Engineers Already Know About Machine Learning Eval datasets, scoring pipelines, and deployment gates for ML systems are not new concepts. Software engineers have been doing this for years under different names.
Open Weight Models and the Separation of Concerns in AI Vertically integrated AI companies collect data, train models, and sell inference. This is a business process error. The organisation that understands the task should own the data that defines it.
The Austrian Case for AI Investment The Chicago school asks what AI allows you to stop paying for. The Austrian school asks what latent structure already exists in your accumulated data, waiting to be read.
Friction Builders and Friction Removers Technical people locate power in the ability to build things. In most organisations, the ability to deliver and the ability to block have considerably more impact.
Capital, Labour, and the Productivity Question Productivity growth has slowed across OECD economies. One explanation: the statistics measure the wrong competition. Capital and labour are not on the same curve.
The Indirection Problem Indirection -- accessing something through a reference rather than directly -- is a C programming concept. It also explains a surprising number of modern frustrations.
Metrics for Business Goals A model with 95% accuracy can destroy business value. How to choose evaluation metrics that reflect what the business actually needs.
Machine Learning Engineer vs Software Engineer: Four Structural Differences Software engineers expect delivery to end at deployment. In ML, that is where the work begins. Four structural differences that change how ML projects are scoped, managed, and judged.
5 Mistakes to Avoid When Building Your First AI Product The traps that derail early AI products are predictable. Five failure modes from years of helping startups build and ship.
How to Brief a Developer on a Feature Without Being Technical A practical guide for business and product owners who need to communicate clearly with AI engineers without speaking code.
Alignment: Semantics in the AI/ML Era LLMs introduced a new kind of problem into software: the model's interpretation of intent can diverge from the developer's. What alignment actually means in practice.
Machine Learning Design Patterns Choosing the right ML design pattern matters more than choosing the right model. Key patterns with applications to marketing and audience intelligence.
The Cost of Machine Learning: A Practical Breakdown Machine learning projects fail budgets more often than they fail technically. One equation, F(X) = Y, breaks down where the cost actually lives across data, model, and labels, and what drives each component up.

Data Strategy (6)

What Is Data Engineering? Two Jobs, One Title Data engineering describes two distinct disciplines that require different skills and produce different failures when confused. Understanding the split is the first step to hiring and structuring a data team that actually works.
What Your Data Already Knows Siloed data is a list. Connected data encodes structure -- communities, gaps, relationships -- that accumulated through ordinary operation and has never been made visible.
SQL Schema Documentation for RAG Pipelines Structured databases contain context that LLMs need to generate accurate queries. How to expose schema information in a way that preserves relationships and business logic.
How to Isolate Bad Training Data Using Cross-Validation Cross-validation is usually framed as a model evaluation technique. It is also the most reliable way to find which training examples are hurting your model.
What Is Data Maturity? Data maturity is a spectrum from ad hoc storage to strategic data infrastructure. What it is, how to assess it, and why it determines what AI is actually feasible.
What Data Do I Need? A Guide for Business and Product Owners A checklist for integrating machine learning into a product without overcommitting data infrastructure before you know what the problem is.

Architecture & Deployment (13)

Serverless on AWS vs Azure: The Real Differences Serverless migrations between AWS and Azure look symmetrical on paper. Lambda's one-function-one-deployment model and Azure's Function App grouping are not the same shape, and that gap costs more time than any other part of the move.
What AWS Bedrock Actually Costs You Bedrock's one-API-many-models pitch does not remove the ordinary difficulty of LLM integration or AWS's own configuration surface. It adds one on top of the other, with no library standing between a team and either.
AI Inference Infrastructure: A Decision Framework How to choose AI inference infrastructure for production: hosted vs. self-hosted, sizing GPUs by VRAM, and serving software like vLLM for handling many concurrent API users. A four-part decision framework.
Open Weights Mean the Tier Is a Deployment Decision, Not a Lock-In With closed AI APIs, your infrastructure tier is permanent. With open-weight models, local, private, and cloud are the same application at different scales. You choose the tier; the code stays the same.
The Private Inference Stack: A Field Guide From raw PyTorch to managed private API services. What each layer of the inference stack does, where the tools come from, and how they relate.
Multi-Tenant Architecture: A Practical Spectrum Multi-tenancy is not a single pattern. It is a spectrum from shared tables to fully separate infrastructure, and the right point on that spectrum depends on what varies between tenants, where your operational complexity budget sits, and what failure looks like.
Bronze, Silver, Gold: A Layered Data Architecture in PostgreSQL The medallion architecture -- bronze for raw ingestion, silver for enriched and validated records, gold for aggregated outputs -- maps cleanly onto PostgreSQL. Each layer has a single concern. The transitions between them are where the interesting engineering lives.
Event-Driven Embeddings in PostgreSQL: LISTEN/NOTIFY and the Async Pattern PostgreSQL's LISTEN/NOTIFY mechanism lets you trigger embedding generation the moment a row is inserted, without polling, without a separate scheduler, and without coupling your embedding service to your database schema.
Private Inference: Running AI Inside Your Own Infrastructure Data egress is the constraint that kills AI projects in regulated sectors. Open-weight models deployed inside a VPC remove the objection before it reaches legal review.
Scaling Databases: Lessons from Facebook's MySQL Journey Facebook's progression from a single database to a sharded architecture contains practical lessons for any system that needs to grow. The constraints that drove each decision.
Platforms vs Pipelines: Engineering for Compounding Returns AWS, Borg, chaos engineering -- successful organisations build platforms rather than optimise pipelines. The distinction determines whether engineering effort compounds.
AI Tooling and the Separation Between Coding and Programming AI-assisted development accelerates a long-running separation between writing code and building systems. Why that distinction matters and what it means for teams.
SaaS Architecture 101 for Non-CTOs LAMP, JAMstack, microservices, serverless -- what these terms actually mean and how to navigate architecture decisions without a technical co-founder.

Infrastructure (7)

Falling Token Prices and the Cost of Inference You Already Own Token prices are falling industry-wide while volume rises faster than the price drops. A measured benchmark of 14 open-weight models on a single GPU shows what that dynamic looks like at the smallest possible scale: 5.6p in electricity for 1,204 requests.
GPU Options for Local AI Inference: VRAM, Software Stacks, and What Actually Works (Mid-2026) Alternatives to standard CUDA-based inference stacks: AMD ROCm, Apple Silicon (Mac Studio, M3 Ultra), Chinese GPU vendors, and which frameworks actually run on non-NVIDIA, non-x86 hardware.
Why Private Inference Is Not Fully Private (Yet) Self-hosted AI solves the data egress problem. It does not solve the computation problem. Fully homomorphic encryption is the missing piece -- and it explains why 'private inference' is currently an architectural approximation rather than a cryptographic guarantee.
Why Private Inference The specific regulatory (NHS, FCA, GDPR) and exposure risks that make private inference necessary for regulated data, distinct from the general architectural case for self-hosting AI models.
Private Inference: Costs Cost claims about self-hosted inference are usually estimates. A full telemetry run across 14 open-weight models on a single consumer GPU replaces the estimate with a number: 5.6 pence in electricity for 1,204 requests.
Deploying Models to AWS Lambda Open-source models on serverless infrastructure cost less than incumbent AI services. A practical guide to deploying on AWS Lambda.
What is Docker? A Deep Dive into Containers The Linux primitives behind Docker containers -- namespaces, cgroups, and how they make containers faster than virtual machines.

Developer Tools (7)

Marigold: A Self-Hosted Inference API for Open-Weight Models An OpenAI-compatible API for open-weight models, running entirely on your own hardware. Open source, installed with pip, airgapped once your models are cached.
Reheat: Intent Clustering for Google Search Console Reheat pulls your Search Console queries, embeds them as vectors, clusters by semantic intent, and surfaces ranked content gaps and opportunities as a report.
How to Build an AI Pipeline That Runs Anywhere: the runfox Backend Model Most workflow libraries couple the pipeline definition to the execution environment. runfox separates them: the same YAML definition runs in-process, against SQLite, or distributed across SQS and DynamoDB, with no changes to the workflow code.
Why JSON Logic Needs JSONPath: Adding the vars Operator Standard JSON Logic resolves single values via dot notation. When rules need to match against lists -- model outputs, tag arrays, multi-value fields -- dot notation is the wrong tool. json-logic-path adds a vars operator backed by JSONPath.
Defining Your AWS API Gateway in Python API Gateway configuration and Python route definitions live in separate files and drift apart. fastapi-aws generates both the AWS integration spec and the public API documentation from a single Python source.
DynamoDB Access Patterns: The Part the Documentation Skips DynamoDB rewards developers who define access patterns before writing schema. boto3 enforces none of that discipline. dynawrap moves the pattern onto the class and lets you swap the backend for PostgreSQL when AWS is not available.
Developing with APIs Without Rate Limits: Introducing APICache Rate limits slow down early-stage development. APICache is an open-source library that caches API responses locally so you can iterate without hitting limits.

Security & Resilience (2)

Are You a Data Controller or a Processor? Two regimes govern analytics data and they test different things. PECR asks whether you stored anything on the device. GDPR asks whether you can identify anyone. A 2025 CJEU judgment changed the answer to the second, and the first is still where most systems fail.
Security Risks in AI Systems with Public Access Publicly deployed AI systems introduce specific vulnerabilities. A breakdown of the key risks and what to do about them.

Perspectives (11)

Generation Speed and the Collapse of Review Delegation Generative video removes the production stage that used to assign taste-authority to a specific role. Review becomes undirected rather than absent, and volume only accelerates a problem that exists at any speed.
Content Negotiation Was a Good Idea HTTP content negotiation was designed to decouple content from form. It stopped at format selection. Generative inference completes the original intention.
The Embedding is the Message AI communication inherited natural language from human communication. That constraint is not a technical necessity.
Adtech Ran This Experiment First The measure-optimise-adjust loop now running through AI first ran through advertising technology. The pathologies are the same. Adtech just got there a decade earlier.
The Fat Protocol Thesis Applied to AI Model weights are the fat layer -- general, beneath everything, available at marginal cost. Workflows are thin clients. Most AI infrastructure is built the wrong way round.
Software Engineering Solved Governance Single source of truth, idempotency, rollback, least privilege -- software engineers formalised governance problems that other institutions still handle by convention.
Content is Dead: The Curation Inversion Generative AI removed the cost constraint that forced editorial discipline. When creation is free and unlimited, the signal collapses. Curation is the new scarcity.
Digital Archaeology: Bitcoin's 21 Million The 21 million bitcoin limit is not in the whitepaper. It derives from a 32-bit integer ceiling, a type migration, and a patch written under pressure after a 184 billion BTC exploit.
Similarities in the Bitcoin and AI Economies Bitcoin and AI both convert electrical energy into discrete computational units via open protocols. The economic structure is closer than it first appears.
PMFOps: Operationalising Product-Market Fit DevOps operationalised code. MLOps operationalised models. PMFOps is the emerging third layer -- treating the audience as a testable, versioned artefact.
The Problem with Faster Horses Foundation models are not fast humans. Treating them as a non-deterministic data store opens more useful questions than the labour displacement framing does.