In a UAE company, an AI automation specialist decides where a language model can safely do customer-facing or back-office work, and builds the checks around it. The hard part is not the prompt. It is the confidence threshold, the PII and duplicate checks, the review queue and the daily sample that tell you the system is still behaving.
Rukhillo Kodirov fits because he has spent two years doing this on live customer traffic, with a rule that nothing unreviewed goes out. He uses Claude Code, ChatGPT, Codex, n8n and Zapier, and he practises on side projects: a 9-agent content pipeline with human approval, and a 6-agent development team that built a tested app. His starting point is always the manual loop and its cost in minutes.
Left: the need. Right: the proof from his own work.
A workflow that classifies customer messages reliably
Classification and routing for 100–200 messages a day, with a 0.90 confidence threshold.
A human in the loop where the model is unsure
Anything below the threshold goes to a short human review queue, with a daily QA sample of 20 routed items.
Checks for PII and duplicates before anything is written
Every classification and draft passes PII and duplicate checks before it reaches the CRM.
A measurable result, not a demo
Agent manual handling fell from 2 hours to 20 minutes a day.
Fluency with current AI tooling
Works with Claude Code, ChatGPT and Codex, plus JavaScript and Node.js for glue code.
Experience directing several agents at once
Runs a 9-agent content pipeline and directed a 6-agent team that built the Falaj finance app, with tests.
Plumbing that other people can maintain
Builds in n8n and Zapier, hands over with a dashboard, a runbook and the before and after number in writing.
He decides where a language model can safely do customer-facing or back-office work and builds the checks around it: a confidence threshold, PII and duplicate checks, a human review queue and a daily QA sample. His focus is customer operations.
With a quality gate: a confidence threshold of 0.90, PII and duplicate checks, a human review queue for anything below the bar and a daily QA sample. Nothing unreviewed goes out.
Claude Code, ChatGPT and Codex for the AI side, and n8n, Zapier, HubSpot and Slack for the workflow around them.
Send me the problem in two lines. I'll reply with what I'd build and what it would change.