Writing
Write-ups on things I actually built, not surveys of things I read about. Each one traces a decision back to the code that implements it, and each says plainly what it does not do. Right now all three are drafts awaiting review.
-
Building Catalogue-Grounded AI Replies for WhatsApp Commerce
A language model that can name any product will eventually name one that does not exist. How the copilot bounds that: retrieval decides, the model only phrases, and any product ID outside the retrieved candidate set is rejected in code rather than discouraged in a prompt.
Builds on the WhatsApp Commerce Copilot.
-
Designing Reliable Human Handoff for Customer-Service AI Agents
Adding a "talk to a human" button is the easy half. Making the transfer survive a restart, a second worker and two messages arriving at once is the half that decides whether a customer ever gets two contradictory replies.
Builds on the WhatsApp Commerce Copilot.
-
Entropy-Aware Data Preservation with PostgreSQL Triggers
Retention logic that no client can skip, because no client performs it — plus a correction I owe the project: the score it computes is a weighted change-density ratio, not Shannon entropy, and there is no logarithm in it anywhere.
Builds on the Entropy-Aware Data Preservation System.