Engineering

AI application architecture

Most of the work in an AI application happens around the model. I care about what information it receives, what it can change, and how the application handles a wrong answer.

Deciding what belongs where

Retrieval, tool use, and generation all sit inside a larger workflow. I try to keep clear boundaries between decisions a model can make and checks that belong in ordinary code.

Context needs the same care. I want to know which sources went into a call and be able to trace an answer back to them.

Planning for mistakes

An incorrect answer can have real effects when it writes to a database or triggers another service. Validation, permissions, and recovery need to be part of the design from the beginning.

Open questions

  • Which decisions are safe to leave to a model, and which must be deterministic?
  • How is the context for a call chosen, recorded, and limited?
  • Where can a generated result write, and what checks it first?