01 / The context
The operating challenge.
Colleagues needed answers from a laboratory information management system containing hundreds of tables. Getting useful information required technical assistance and an understanding of business meaning that was not always apparent from the database schema.
The opportunity was to make routine research and operational questions easier to answer while preserving individual access permissions, traceability, and the integrity of the source data.
02 / My contribution
What I was accountable for.
I presented the investment case to the CEO, secured approval for a two-person, two-month build, and architected the product and its controlled query workflow.
I worked with the team through approximately two months of iteration on prompting, workflow logic, local-model performance, and accuracy. Cross-department testing with approximately 50 colleagues helped expose domain questions that schema names alone could not resolve.
03 / Key decisions
The choices behind the work.
Use the capacity already in place
The investment case used existing locally hosted models and Kubernetes capacity. The team could concentrate on the workflow, quality, and access controls.
Make permissions part of the product
Read-only database accounts, individual identities, row-level security, and audit logs constrained access. Query syntax and executability were validated before execution.
Evaluate business meaning
Testing included questions whose interpretation depended on domain knowledge. Iteration addressed the meaning of questions as well as whether a query could run.
How the workflow connects
- 01Ask
A research or operational question
- 02Validate
Query syntax and executability
- 03Retrieve
Read-only, permission-controlled access
- 04Review
A result with an auditable query trail
04 / What changed
Results in the operating environment.
After launch, Ask-LIMS served approximately 300 monthly active users and 10,000 research and operational-analytics queries per month.
Typical retrieval fell from one or two hours to approximately one minute. The product supported research and statistics; its outputs were not intended for direct clinical decisions.
05 / The lesson
Useful enterprise AI depends on domain understanding, permissions, and evaluation as much as model capability. Those constraints belong in the product architecture from the beginning.
Figures are approximate where indicated and reflect the period described.