RAG Quality as a Distribution, Not a Single Score
A retrieval-augmented answer can look excellent in a demonstration and fail for a reader who asks the same question differently. It can cite a real document while relying on the wrong version.
THE AUTHOR
Marcin Borecki
Developer, entrepreneur and business angel. Writing about AI, software, data and the evidence behind technical decisions.
About the author →THE READING ROOM
Writing & investigations.
The Command Completed, the Goal Did Not: SQL Maintenance Postconditions
A real SQL Server fixture successfully refreshes the wrong statistic, rejects the resulting state, and verifies the correct repair. Command success is not the maintenance goal.
After the Restart: Durable Replay and External Effects in a .NET Agent
A real child process exits after a local receiver commit. Reconciliation preserves one effect while keeping unestablished outcomes explicitly unknown.
AI Changes the Cost of Coordination, Not Just Creation
When creation gets cheaper, work moves into selection, review, integration and operation. The economic unit should be a maintained outcome.
One Question, Several Queries: Put Typed Intent Between Language and SQL
Typed intent separates language proposals from executable SQL, trusted tenant scope and independently checked aggregate results.
When Monitoring Becomes the Workload
A monitoring system can correctly report that a database is struggling while helping to make it struggle. That is an uncomfortable property of observability: the observer participates in the system it describes.
A Verified Backup Is Not Yet a Recovered Application
A real checksummed backup is verified, restored and checked against exact transactions, then the application refuses readiness until its external dependency agrees.
Conversation State Is Not Application State: A .NET AI Reference Repository
A real ChatClientAgent read-tool roundtrip shows why a fresh conversation can forget history while SQLite preserves the current domain record.
What a Performance Diagnostic Should Deliver
A useful diagnostic produces a reproducible finding and a decision, not just a large dashboard or a list of suggestions.
Deployment, Runtime and Operator: Three Identities in One Azure Service
Separate deployment, runtime and operator authority with actual local JWT checks, then inspect an explicitly undeployed Azure identity mapping.
A Career Built on Judgment, Not Syntax Recall
Engineering judgment is built through specific decisions and feedback. A career strategy should develop that ability without treating technical fluency as disposable.
Vector Search Under Pressure: Distance, Candidates and Index Updates
An exact baseline and a materialized sign-bit candidate index expose where search quality is lost, why update identity matters and why fewer candidates did not always run faster.