Our summary
An AI assistant can save two records for one person, or wrongly join records for different people with similar names. Ankush describes how Construct uses Jev to compare uncertain mentions of people and projects with its saved memory. Clear matches are handled by ordinary software first.
For the remaining cases, software gathers a small list of candidates. Jev compares them together, including a no-match option. Construct joins records only when the chosen match clears its 0.7 confidence threshold. If Jev times out or returns a malformed answer, the previous AI judge takes over.
The threshold came from seven probes, not a broad benchmark, so another team needs its own examples. Only entity matching is live; task-completion and email checks are disabled. Construct also lacks mandatory approval before every external action, so sending mail or moving money still needs supervision.
Key takeaways
- Resolve clear matches in ordinary software first, then give Jev only the uncertain candidates and an explicit no-match option.
- Choose a confidence threshold from your own examples and the cost of mistakes; seven probes do not establish general accuracy.
- Keep a fallback for missing or malformed answers, record the model version, and do not treat confidence as permission for risky actions.