Build the platform first. Make evaluation part of engineering.
Two connected leadership choices: create the reusable technology foundation before the visible product, and make evaluation a property of production systems rather than a final review.
- Organization
- Jewelers Mutual
- Period
- 2024 — 2026
- Record
- Decisions 02 + 03
System connections
- Architecture & platform
- Product & delivery
- Evaluation
- Business outcome
Context
The modernization mandate created an immediate sequencing question: ship a visible product quickly, or first build the foundation that later products and workflows would need.
Production AI introduced a second constraint. In a regulated environment, a workflow that is right most of the time is not sufficient; explainability, evaluation, and responsible-use controls have to be part of how the system operates.
Decision
I architected and delivered an event-driven, serverless AWS platform first, alongside standards for API governance, observability, DevSecOps, and event-driven architecture. Anything Promise then became the first product delivered on that foundation.
For production agentic work, evaluation and ethical controls were treated as engineering requirements. They were designed into the system rather than added as a review before launch.
Tradeoff
Building the platform first delayed the first visible product, but reduced the risk of producing a one-off solution that the next product could not reuse.
Making evaluation continuous required more discipline throughout delivery. The alternative was a faster path to launch with less confidence that the system would remain trustworthy in production.
Outcome
~200 Hours of manual work removed each month by a production agentic workflow
Anything Promise validated the new platform as the first product delivered on it and accelerated time to market.
In separate production AI work, an agentic workflow eliminated approximately 200 hours of manual work each month without exposing confidential workflow or customer details.
Evidence
The public claims in this case study are limited to approved wording and supported by the following published artifact.