Written from things we operate.
Engineering positions are stated as engineering opinion. Where a piece covers something we intend to measure rather than something we have measured, it says so — no performance figures appear here that we have not produced and published.
Engineering articles
- 01Field note
What breaks first when an agent goes to production
A demo agent and a production agent are different systems wearing the same name. The differences are almost never in the prompting.
- 02Deep dive
Multi-tenant RAG: enforcing scope so tenant A cannot read tenant B
If your product is multi-tenant, retrieval is a data-boundary control. Build it like one rather than trusting every caller to remember.
- 03Deep dive
Regression evals in CI: catching the prompt change that broke three other things
Every generative feature has a moment where it silently gets worse. Without evals gating the build, the detection mechanism is a customer complaint.
- 04Deep dive
Cost-aware model routing: spending less without getting worse
Most steps in an agent workflow do not need a frontier model. Routing is usually the largest available saving and the least disruptive to make.
- 05Deep dive
Supervisor or swarm: choosing an agent topology you can test
Topology decides how testable, debuggable and expensive a system will be. It is worth more deliberation than it usually gets.
- 06Field note
Hybrid retrieval: why dense-only search keeps missing the obvious
Semantic search is bad at exact tokens. Most 'the AI gives bad answers' complaints are retrieval complaints wearing a costume.
- 07Methodology
Writing acceptance criteria for work whose output is non-deterministic
Fixed-price AI engineering is only workable if 'done' can be written down. Here is how to write it when the output varies by run.
- 08Field note
The gap nobody scopes between a working prototype and a shipped feature
The prototype is done in a week and the feature takes three months. That ratio is not incompetence; it is the work being different.
- 09Deep dive
Tool-call failure modes, and designing agents that recover
Tools fail in more ways than they succeed. The recovery path deserves as much design attention as the happy path.
- 10Methodology
Evaluating retrieval when you have no labelled data
Waiting for a perfect golden set means never measuring anything. Here is how to bootstrap a baseline in an afternoon.
- 11Field note
Latency budgets: what users tolerate when they know a machine is thinking
Agent systems are slow by construction. Tolerance depends far more on framing and feedback than on raw seconds.
- 12Field note
Operating a portfolio of AI products as one person
Running several products alone changes which engineering decisions are correct. Anything requiring routine attention is a liability.
Creative articles
- 01Methodology
Is AI-generated UGC legal? What the FTC and the EU AI Act require
Short answer: yes, with disclosure obligations that are stricter than most advertisers currently assume — and an EU deadline that has arrived.
- 02Methodology
Disclosing AI-generated ads without wrecking the ad
Disclosure is required. It is also a craft problem, and treating it as one produces better results than treating it as a legal tax.
- 03Methodology
Synthetic creator or real creator: what actually differs
The comparison is usually framed as cost. The more useful differences are trust, claim safety, iteration speed and where each one breaks.
- 04Methodology
The economics of AI creative: what actually costs money
Generation is nearly free. Selection, direction and revision are not — and they are where the real budget goes.
- 05Field note
Hooks: the first two seconds decide whether the rest was worth making
Cheap generation makes hook testing affordable for the first time. Most teams still write one and hope.
- 06Field note
App-launch creative: what a cold audience needs in fifteen seconds
A cold viewer does not know your category, your product or your name. Most launch creative assumes at least one of those.
- 07Methodology
Creative testing at scale without burning the budget
Volume without a testing structure is just expensive noise. The structure matters more than the volume.
- 08Field note
What AI UGC gets wrong, and how to fix it in production
The failures are consistent and mostly avoidable. Almost none of them are model limitations.
- 09Deep dive
Motion product ads from stills: keeping the product itself consistent
The giveaway is drift — the product subtly changing between shots. Asset-lock before motion is the fix.
- 10Methodology
Does labelling an AI ad hurt performance? The question worth measuring properly
Everyone has an opinion. Almost nobody has run it cleanly — and the opinion is being used to justify non-compliance.
- 11Methodology
Rights, likeness and voice: the consent stack for synthetic talent
A generated person is not a free person. Here is the paperwork that has to exist before a campaign ships.
- 12Field note
Why the same ad works on an app and fails on a SaaS product
Different decisions, different time horizons, different proof requirements. The creative has to reflect that.