01AI LAYER
INTERPRET
- Understand context
- Extract meaning
- Compare evidence
- Generate proposals
Probabilistic judgement
ENGINEERING APPROACH
From ambiguityto a controlled system.
Practical AI without theatre.
INPUT → INTERPRETATION → CONTROL → ACTION
I use AI where interpretation matters, deterministic software where control matters, and human review where the cost of being wrong is high.
DOCUMENT
IMAGE
WEBSITE
AUDIO
BUSINESS DATA
SYSTEM CONTROLLED
SYSTEM STATECONTROLLED
AI → SOFTWARE → HUMAN
The Architecture
of Control
Reliable systems assign each responsibility to the right layer.
02DETERMINISTIC CODE
CONTROL
- Validate outputs
- Enforce permissions
- Route workflow state
- Apply limits and policy
- Reject invalid results
Predictable enforcement
03HUMAN REVIEW
DECIDE
- Review ambiguity
- Approve high-impact actions
- Resolve missing context
- Remain accountable
Final responsibility
AI INTERPRETSSOFTWARE CONTROLSHUMANS DECIDE
The goal is not maximum autonomy.
It is appropriate control.
PROBLEM → SHIPPED SYSTEM
The Engineering
Process
A repeatable path from uncertain input to a shipped system.
DEFINE
Define the real problem
Solve the underlying uncertainty, not only the visible request.
MAP
Map inputs and constraints
Identify data sources, edge cases, limits and dependencies.
SEPARATE
Separate interpretation from control
AI proposes. Deterministic software enforces.
BUILD
Build the smallest complete system
Connect the input to a real user action before optimising.
VALIDATE
Validate with real behaviour
Test messy inputs, failures and actual user decisions.
SHIP
Ship and iterate
Observe the system, measure failure and tighten constraints.
The system can be complex.
The next step should not be.
Complexity belongs in the architecture—not in the user's action.
VIEW SELECTED WORK