// Engineering notes
AI in Practice
Practical lessons from implementation and architecture investigation. Each anonymized note explains the problem, decisions and verification.
Implementation note
When valid JSON contains an unusable AI response
A response can pass structural validation and still contain text that should never reach the user.
Read the note →Architecture investigation
A vector search module is only part of conversation memory
Retrieval becomes useful when information is saved, scoped to the right person and used in the live workflow.
Read the note →Implementation note
Changing an AI model reaches beyond the API call
Model selection, stored settings and cost calculations need to agree after a migration.
Read the note →