How we think about AI systems
A short manifesto from our founder. It is what we believe after designing and shipping AI systems for lenders, insurers, schools, retailers and enterprises.
1. The model is the smallest part
In every production AI system we have built, the model accounted for a small fraction of the code, the cost and the risk. Retrieval, data contracts, evaluation, caching, permissions, fallbacks and monitoring made up the rest. Teams that design the model first and the system later end up rebuilding the system.
2. If you can't measure it, you don't have it
"The assistant is pretty good" is not a state a system can be in. It has a recall at k, a faithfulness score, a p95 latency and a cost per answer, or it has an unknown quality that will be discovered by users. We build the golden set before the prompt.
3. Autonomy is earned, one action at a time
Agents should start by observing, then recommending, then acting with review, then acting alone, per action type, based on measured agreement with the humans who did the job before. A system that begins fully autonomous begins with the trust it has not yet earned.
4. The simplest model that meets the target
Gradient-boosted trees for tables. Rules where rules suffice. A small model behind a cache before a large model on every request. Novelty is a cost, and we spend it only where it buys something the client can measure.
5. Data residency is an architecture decision, not a legal footnote
Whether data may leave your environment shapes the vector database, the model choice and the deployment topology. We ask on day one, and we have shipped fully self-hosted systems when the answer was no.
6. Every system should survive its builders
Architecture decision records, runbooks, tests and documentation are how a system outlives the people who built it. We deliver them as part of the work, not as a change request, because we expect your engineers to own what we hand over.
7. The architect reviews the build
Design intent is lost in handoffs. The person who designed the system should see the implementation, review the pull requests that matter and sign off the milestones. That is how our engagements are structured.
Anas, Founder & CTO
Put this into practice
Our AI readiness checklist turns these principles into twenty questions.
Agree? Disagree?
Either way, we'd like to hear about the system you're designing.