Architecture review.
When one decision is too important to get wrong.
Most senior architecture decisions in AI get made by people who have never put a comparable system into production. A vendor's solutions architect. A consultant who has read the documentation. A senior engineer who has built one prototype and one production system. That is not enough.
An architecture review brings someone who has shipped this exact class of system before, looks at what you are proposing or what you have already built, and tells you what they would change. In writing. With named risks tied to specific layers of the stack.
When this fits
- A vendor pitched an architecture and you want a second read before you sign.
- Someone on your team built a prototype and you are trying to decide whether to ship it.
- You suspect a system is drifting on cost, latency, or quality, and you cannot prove it.
- An RFC, design doc, or proposal is on your desk and the decision is hard to reverse.
- A board or audit committee has asked you to validate an AI design choice.
When it does not
- There is no specific artifact yet. Start with the discovery phase of a project build.
- You need someone embedded in the work over time. See engineering leadership.
- You want an opinion, not a review. We will not price an engagement without something to evaluate.
- Step 01
Scope call
A 30-minute call to confirm what is being reviewed, what you are optimizing for, and what good looks like. If the scope cannot be defined, we say so and we do not take the engagement.
- Step 02
Review
Three to five business days reading the artifact: code, traces, evals, vendor docs, RFCs, architecture diagrams. We follow up in writing where we need more.
- Step 03
Written assessment
A document with what we would keep, what we would change, what we would watch for, and the specific stack layer each finding lives in. Specific. Ranked by risk.
- Step 04
Optional walkthrough
A 60-minute call with your team to walk the assessment, answer questions, and resolve disagreements. Included on request.
- A written assessment from a senior operator who has shipped this class of system.
- Findings tied to specific layers of the stack: architecture, correctness, routing, integration, observability.
- A short list of things to change before production.
- A short list of things to watch once the system is running.
- Owns Architecture
Henrik Soerensen
Principal AI Solutions Architect
Vice President of Engineering at MGM+ and Amazon Prime Video. Led a 34-person organization at 99.999% uptime.
Read profile ↗ - Owns Correctness
Jay McCarthy, PhD
Chief Scientist
PhD in Computer Science, Brown. Co-creator of the Racket programming language. 32 peer-reviewed publications.
Read profile ↗
From $1,200 for a single-artifact review. Scoped per engagement. Multi-component reviews and quarterly check-ins are quoted on request.
Tell us what you are looking at. We will tell you whether this is the right engagement, a different one, or none of them.
Direct line.
info@swenor.us