Writing

Notes from real projects on AI agents, data platforms and system architecture. Written in plain language.

Engraving of a horse's head wearing a harness collar and bridle
FeaturedAI Engineering

We didn't need a better model. We needed a better harness.

When an AI agent gives confident, wrong answers, the model is rarely the problem. Here is how I fix the harness around it and improve it every week.

From building AI agents for clients · 6 min read
An old engraving of the Louvre in 1677, with part of the building wrapped in wooden scaffolding
Architecture

Low-code at scale: fix it, rebuild it, or add AI?

Your low-code app worked at first. Now releases are slow and the team lives in Excel. Three paths forward, and how to pick the first one.

For teams running a low-code platform at scale · 6 min read
Black and white photo of a wooden library card catalogue with rows of labelled drawers and metal handles
Data Engineering

One set of numbers: scaling a VC firm's data platform on Snowflake

How I approach a VC firm's data platform on Snowflake: Airflow, dbt, metric owners, access control, cost control and data ready for AI agents.

From running Snowflake for a VC firm · 6 min read
An old hand-drawn lamp sketch on the left, an arrow, and a clean engraved lamp patent drawing on the right
Visual AI

From a rough sketch to a clean drawing: how I use AI on messy visual input

A practical pipeline to turn hand sketches, whiteboard photos and scans into clean, editable drawings: clean, find shapes, style, review.

On turning rough visual input into structured drawings · 6 min read
Etching of a sailing ship with full sails on open water
Career

9 years, 6 companies: what makes a data or AI project actually ship

Lessons from nine years of data and AI work, from strategy to code: why some projects reach real users and others stall.

From nine years across six companies · 6 min read
An old iron padlock with a key in it
Security

Prompt injection, step by step: how a harmless request leaks your data

A step-by-step demo of indirect prompt injection, why the model is not a security boundary, and the defences that contain it.

From building agents that read documents and call tools · 6 min read
The Oberbaum Bridge in Berlin, a red brick double-deck bridge with towers crossing the River Spree
Data Engineering

Moving to a lakehouse without stopping the business

A five-phase method to move old warehouse pipelines to a lakehouse while reports keep going out: map, land, rebuild, compare, switch.

From migrating legacy pipelines to Databricks on AWS · 5 min read
tool servicesSlackAirflowAgentCloud RunWeb fetchSearchWarehouse
Architecture

Running AI agents on Google Cloud: keep the tools outside the agent

An architecture for company agents that people use from Slack and Airflow, that query the warehouse, and that stay safe to change.

From building an internal agent platform on Google Cloud · 2 min read
where most answers go wrongDocsowner, dateChunksclean splitsIndex+ metadataModelanswersfix the data before you blame the model
AI Engineering

RAG in production is mostly a data problem

Most bad answers from a RAG system start before the model is even called. Where they come from and how to measure them.

From putting retrieval over company documents into production · 2 min read
prompt v1prompt v2prompt v3Test setfixed casesv162%v278%v391%
AI Engineering

Prompts are code. Treat them that way.

Small wording changes can move accuracy a lot. Version prompts, test them on a fixed set, and let real code do the maths.

From shipping LLM features to production · 2 min read
Salesowns ordersServiceowns diagnosticsProductowns usageFinanceowns invoicesMarketingowns campaignsSupportowns ticketsShared platformstorage, catalog, access
Architecture

Data mesh in practice: start with ownership, not tools

Data mesh is mostly an agreement between teams. The simple version that worked for me across six teams.

From moving data ownership to domain teams · 2 min read
PostgresAPIs, filesS3rawSnowflakeclean modelsBIdashboardsAirflow schedules and retries
Data Engineering

A small, solid data warehouse on AWS

How I set up S3, Airflow and Snowflake so a growing company could trust its numbers, and what I would still do the same way.

From building the data warehouse for a fintech team · 2 min read