Skip to content

Latest test, 20 Oct 2026: Haiku 4.5 vs Haiku 5.5 vs Mercury 3. See the scoreboard →

Andrew Deighan

src/data/facts.json · 51 entries

Fact file

Every claim on this site, in one place. Site copy is written from this file, and the role-match agent may only state things that trace back to an entry here. That includes the scope: what the work does not cover.

p1Andrew Deighan. British and Irish citizen with full UK and EU right to work, relocating to the UK. Open to roles in the UK, Ireland, the EU and remote. Available from mid-October 2026.
p2Co-founder and Product Lead (CEO) of AtlasOra, June 2024 to October 2026. Two-sided vacation rental marketplace. Led a team of thirteen.
p3Founder of DUDsJobs, an AI job-hunting service, live with paying users since September 2026.
p4Head of Marketing at Qubic, an AI-focused layer-1 network, August 2024 to May 2025.
p5Nine years in secondary education across the UK, Thailand, New Zealand and Jordan, finishing as Assistant Headteacher on a school senior leadership team. Trained as a British Schools Overseas inspector.
p6MSc in Management (FinTech & Digital Banking): taught modules complete with an 84% average, dissertation pending. PGCE. BSc Pharmaceutical Science.
p7Builds with AI coding tools, including Claude Code. Not a trained software engineer.
p8Served in the RAF from 2002 to 2007 as an electronic technician on radar systems, holding SC clearance.

Host acquisition agent pipeline

project page →
h1An orchestrated pipeline of 11 working agents and rule-based gates. Two further agents are registered but not yet built out, and are not counted.
h2It worked through all 117 property managers on the Costa del Sol: finding decision-makers, researching each company and drafting personalised outreach.
h3Around 20 managers agreed to list about 1,000 properties in under three weeks. These are agreements to list, not live listings.
h4Email outreach got about 2 replies. A phase running email and WhatsApp side by side, each manager on one channel, then a full switch to WhatsApp got about 45 replies (between 40 and 50).
h5Models are tiered by task: Claude Haiku to sort, Sonnet to research, Opus to write.
h6Every draft passes an automated tone check (rules, then a model judgement). Failed drafts are rewritten up to three times.
h7A rule-based compliance gate checks throttling, GDPR lawful basis and consent before every send, and writes an audit record.
h8Nothing is sent without a person approving it. The agent that handles replies never replies by itself.
h9Failure handling: automatic fallback to a smaller model, an alert when fallback passes 10% in an hour, retries with backoff, a daily spend cap and kill switches by agent, channel and region.
h10Several agents use no model at all (prospect finding, sending, compliance, follow-up templates). That is deliberate.

AI job matching engine

project page →
j1A five-model pipeline that routes each task to the cheapest model able to do it well: a low-cost gate model (Jev), Claude Haiku for extraction, Sonnet for matching, Opus for CV parsing and writing, with DeepSeek as a second provider.
j2Rules and the gate settle 82% of job assessments. The mid-tier model reads 18%.
j3The gate rejects 47% of candidates in under half a second.
j4Median AI cost is $0.09 per search run.
j5Failover to DeepSeek is automatic when the primary provider errors, times out or rate-limits, after three retries. It never triggers on a bad request or a refusal. Anything DeepSeek scores is re-scored by the primary model on the next run.
j6Each user has a hard monthly AI budget. When it is reached, matching continues on rule scores alone.
j7Live with paying users since September 2026. 74 of 78 production runs succeeded; all four failures were in the launch cut-over.
j8The job data feed costs more than the AI.

Decision Intelligence

project page →
d1A guest-facing feature that scores each home as a percentage match against the guest's priorities, using measured distances, noise, sunlight, listing details and reviews.
d2Runs only when the guest asks. It never books and never reorders results.
d3AI judgement is labelled separately from measured fact.
d4Multi-model design: Jev, TypeSafe's decision model, makes the scoring decisions, and a language model reads the guest's request and explains the result. In production within weeks of Jev's public launch.
d5Shipped in stages: a hand-written evaluation set with a 100% pass bar, then a two-day shadow run against a standard LLM pipeline, then a two-day canary over several hundred live requests.
d6The launch gate was zero false claims: never calling a home close to the beach when it is not.
d7The model's instructions were iterated against the evaluation set to fix place ambiguity, such as an Irish bar matching an Irish cultural centre.
d8If a model fails, it falls back to measured facts alone. This kept the feature working through a model outage in September 2026.
d9Guest outcome data (usage, bookings) had not been collected by handover in October 2026.

AI Help and host assistant

project page →
a1A guest help assistant grounded in a 497-entry knowledge base. For each question it retrieves the closest entries and answers from them.
a2It is barred from stating numbers, policies, dates or features that are not in its sources, and points unresolved issues to support.
a3This was built after the assistant was found inventing facts such as dates.
a4A separate host assistant drafts profile and listing text and can change only a fixed set of fields.
a5The safeguards on AI Help are instructions and grounding. There is no automated checker on its answers.

Wider product and payments work

on the home page →
w1Took AtlasOra from zero to public launch in September 2026: a 198-feature requirements document and 98 features shipped with a CTO and four engineers.
w2In-person discovery with property managers across Malaga and Marbella.
w3Designed and specified trigger-released escrow contracts on Base. Guest funds were held in EURC and released at check-in, or automatically 24 hours later, with dispute freeze and admin settlement. It ran in production and is currently switched off for cost reasons.
w4Automatic cross-border host payouts by Revolut bank transfer once the on-chain release was recorded, so hosts never touched crypto.
w516 live API integrations with property management systems (a 17th built and ready), plus a payment provider and KYC.
w6Led a 13-person team and hired and developed three product designers.
w7Shipped on four model providers: Anthropic, OpenAI, DeepSeek and Jev.
q1Worked directly with the project's AI scientists, David Vivancos and Dr Jose Sanchez.
q2Hosted their public technical interviews and supported the launch of their research paper on Aigarth.

Scope

2 entries
s1Shapes model behaviour through instructions, grounding, evaluation sets and graders. Has not fine-tuned model weights.
s2Works in English.