Skip to content
Andrew Deighan

Andrew Deighan

AI Product Lead

I design, build and ship AI products: agents, evals and multi-model systems.

Available mid-October 2026 · Based in Dubai, open to the UK, Ireland, EU and remote.

Work

Experiments

About

Co-founder and Product Lead at AtlasOra, a two-sided vacation rental marketplace: from zero to public launch in September 2026, with a CTO and four engineers, leading a team of thirteen.

Before that, nine years in secondary schools across the UK, Thailand, New Zealand and Jordan, finishing as Assistant Headteacher.

Served in the RAF from 2002 to 2007 as an electronic technician on radar systems, holding SC clearance.

Every claim here traces to the fact file, gaps included.

Match me to your role

Paste a job description

An AI agent reads the role against my record and says where I fit, with evidence, and where I don’t. Nothing runs until you press the button.

Paste a job description
Try an example:

0 / 15,000

What you paste is sent to an AI model to produce the answer and is logged without your name or IP address, so I can see what roles people are hiring for. Don’t paste anything confidential.

How this worksexpand
Model
Claude Haiku 4.5 (claude-haiku-4-5), called only from the server. Nothing runs until you press the button.
Source
It answers only from a fixed file of 53 facts, which you canread in full, gaps included. Up to 15,000 characters are used; longer text is trimmed, not refused. Equal opportunity statements, privacy notices, cookie text and application form fields are removed before the model sees it. Your text is treated as an untrusted description of a role, and any instructions in it are ignored.
Second check
Rules in code drop any claim that cites no fact, cites a fact that doesn’t exist, runs too long, or uses a number its facts don’t contain. Then a second model call checks each remaining claim against only the facts it cites and removes anything it can’t trace.
Limits
5 runs per visitor per day. The agent stops for the day once estimated spend reaches$2.
Failure
If the model is slow, fails, or its answer doesn’t survive the checks, you get a short static summary instead. Nothing else on this site depends on the agent.
Tested
Across the last 3 full runs of the evaluation set (5 October 2026), 79 / 84 cases passed, with no invented facts in any of them. The cases include roles I fit, roles I don’t, text that isn’t a job description and prompt-injection attempts.

Contact

email
[email protected]
linkedin
linkedin.com/in/andrewdeighan
cv
Download CV pdf
available
Mid-October 2026 · Dubai, open to the UK, Ireland, EU and remote