Andrew Deighan · AI product
claude-haiku-4-5), called only from the server. Nothing runs until you press the button.| Measure | Value | Fact |
|---|---|---|
| Properties agreed to list in under three weeksHost acquisition agent pipeline | ~1,000 | h3 |
| Reply rate after one measured channel changeHost acquisition agent pipeline | 1% → 25% | h4 |
| Median AI cost per search runAI job matching engine | $0.09 | j4 |
| properties agreed to list in under three weeks | ~1,000 |
| reply rate after one measured channel change | 1% → 25% |
| median AI cost per search run | $0.09 |
11 working agents and rule-based gates that researched every property manager on the Costa del Sol and drafted outreach for a person to approve.
A five-model pipeline that sends each task to the cheapest model able to do it well. Rules and a cheap gate settle 82% of assessments.
A guest-facing feature that weighs homes against a guest’s priorities, labels its judgement separately from measured fact, and falls back to facts when the model fails.
A guest help assistant that answers only from a 497-entry knowledge base, and a host assistant that can change only a fixed set of fields.
| # | Rule | Example from the work | Fact |
|---|---|---|---|
| 1 | Rules first, models where judgement is needed.Rules and a cheap gate settle 82% of job assessments in DUDsJobs. A model reads the other 18%. | Rules and a cheap gate settle 82% of job assessments in DUDsJobs. A model reads the other 18%. see it: Rules first, models where judgement is needed. | j2 |
| 2 | A person approves anything that leaves the building.No outreach from the host acquisition pipeline is sent without a person approving it. | No outreach from the host acquisition pipeline is sent without a person approving it. see it: A person approves anything that leaves the building. | h8 |
| 3 | Check the output, then check the checker.Every outreach draft passes a tone check and is rewritten up to three times. An alert fires when model fallback passes 10% in an hour. | Every outreach draft passes a tone check and is rewritten up to three times. An alert fires when model fallback passes 10% in an hour. see it: Check the output, then check the checker. | h6h9 |
| 4 | Design the failure before the feature.Decision Intelligence falls back to measured facts when the model fails. It kept working through a model outage in September 2026. | Decision Intelligence falls back to measured facts when the model fails. It kept working through a model outage in September 2026. see it: Design the failure before the feature. | d4 |
| 5 | Measure, then change course.Email outreach measured a 1% reply rate. Moving to WhatsApp took replies to 25%. | Email outreach measured a 1% reply rate. Moving to WhatsApp took replies to 25%. see it: Measure, then change course. | h4 |
| Measure | Value | Fact |
|---|---|---|
| Public launchFrom zero, as co-founder and Product Lead (CEO) | Sep 2026 | w1 |
| Features shippedFrom a 198-feature requirements document, with a CTO and four engineers | 98 | w1 |
| Team led | 13 | w6 |
| Product designers hired and developed | 3 | w6 |
| Live integrationsIncluding a payment provider, KYC and seven property management systems | 11 | w5 |
| Model providers shipped onAnthropic, OpenAI, DeepSeek and Jev | 4 | w7 |
| Escrow on BaseCurrently switched off for cost reasons | ran in production | w3 |
I designed and specified trigger-released escrow contracts. 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.
Automatic cross-border host payouts by Revolut bank transfer once the on-chain release was recorded, so hosts never touched crypto.
As Head of Marketing at Qubic, an AI-focused layer-1 network, I worked directly with the project’s AI scientists, David Vivancos and Dr Jose Sanchez.
I hosted their public technical interviews and supported the launch of their research paper on Aigarth.