Essays on AI-assisted development governance, engineering systems, product strategy, and digital transformation.
Featured
AI Governance
AI coding tools write at machine speed. Humans review at human speed. That gap is widening every sprint -- and most engineering teams haven't noticed yet.
Read essay →All Essays
The problem with AI coding agents isn't intelligence -- it's memory. Every session starts cold. Mneme HQ is my answer to architectural drift in long-running AI-assisted projects.
AI coding tools write at machine speed. Humans review at human speed. That gap is widening every sprint -- and most engineering teams haven't noticed yet.
Attribution is framed as a modelling problem. It is not. Until data unification is solved, the model you choose is the least of your problems -- it is running on unreliable inputs.
How Mneme HQ stores decisions as version-controlled YAML, injects them into AI context at generation time, and enforces them before code reaches review -- with a walkthrough of each layer.
Data lakes were built for batch analytics. AI has different requirements: freshness, schema consistency, governance, and retrieval precision. Most organisations discover this gap only after trying to build on it.
Post-generation review is the wrong place to catch architectural violations. Governance must move to generation time -- enforcing constraints before AI-generated code reaches human review.
The problems showing up in AI systems today -- data quality failures, measurement confusion, governance debt -- are the same problems analytics teams have been navigating for decades. The tools are different. The patterns are not.
Prompting an AI to follow your architecture is not the same as enforcing it. One is a request. The other is a constraint. The difference matters more than most engineering teams have reckoned with.
AI coding tools have been adopted at scale without the governance frameworks to match. The result is a structural gap between generation velocity and the organisational capacity to maintain coherent, auditable codebases.
Open source isn't just a distribution strategy. For developer tools, it's the fastest form of market validation available -- and the only one that produces the kind of trust that enterprise adoption requires.
Concrete numbers from 90 days of decision tracking on real projects: how often AI agents drift, what kinds of decisions get violated most, and where the governance layer pays for itself.
Two ventures, one thesis. Mneme HQ brings structured memory to AI-assisted development; CannabisDealsUS brings structured pricing to a 23-state legal patchwork. Both are bets that the value lives in the data layer.
A walk-through of the Mneme MCP server: how it exposes a project's architectural decisions to Claude Code, Cursor, and any other MCP-compatible agent -- so the agent fetches its own constraints instead of waiting for the human to paste them in.
What it actually takes to keep a multi-state cannabis pricing dataset accurate: schema drift, address normalisation, regulatory variance, and the unglamorous work that decides whether your data is useful or just present.
AI coding agents honour an architectural constraint at the start of a session and quietly violate it later. The failure is not intelligence. It is memory.
Most dashboards go unread because they start from available data. Invert it: start from the decision, name its owner, its cadence, and its trigger.
AI agent reliability is not a model property. It is an engineering discipline data engineers already know from ETL: idempotency, validation gates, observability.
An honest postmortem of shipping Mneme HQ's MCP server: scope mismatches, protocol version skew, stale context injection, and schema validation failures.
The next step up the analytics maturity curve is not a better dashboard. It is a shift from reporting what happened to reasoning about why and what to do next.
Essays cross-posted to LinkedIn with discussion. New pieces every 3 days.
Follow on LinkedIn
© Theo Valmis