Internal workflow automation

The problem

Describe the manual process: who did it, how often, and how long it took.

Approach

The existing manual process mapped first, including the exception cases, which is where automation usually breaks.
The deterministic majority of the workflow automated with no model involved at all, which shrank both the problem and the running cost.
A single AI classification step added where a human judgement call was genuinely required, with a confidence threshold below which the item went to a review queue.
Every run traced end to end: inputs, tool calls, results, duration and cost.
Run in shadow mode, proposing, with a person approving, for an agreed period before any autonomous operation.

More work

Backend & infrastructure

Multi-tenant Laravel platform

A Laravel application serving multiple client organisations from one codebase, with per-tenant data isolation and queued background processing.

LaravelMySQLRedisHorizon +2

Server engineering

cPanel hosting fleet migration

Moving a set of cPanel accounts onto new infrastructure, websites, databases, mailboxes and DNS, with the cutover planned so downtime stayed in minutes.

WHMcPanelIMAPSyncrsync +2

AI engineering

Retrieval-grounded support assistant

A support assistant answering from the organisation's own documentation, with citations on every answer and a clean handover when confidence was low.

Claude APIPythonFastAPIpgvector +2