
Every technical decision, for a decade and a half.
For a decade and a half I was the technical director of StudentTenant.com, a leading online student-lettings platform in the UK: product, infrastructure, other people's personal data, and the team. As AI changed what software has to defend against, my focus moved to AI and security. I stayed on as technical advisor, and the platform still runs on the foundations I set. These are some of the challenges along the way.
StudentTenant.com
Measured in production, not in a demo.
Challenges over the years
Erasure, built before the law applied
Six weeks before GDPR applied, the platform already had a right-to-be-forgottenRight to erasureUnder GDPR (Article 17) a person can ask an organisation to delete their personal data. The organisation must do it without undue delay, unless the law requires it to keep something.Learn more → process in production: a request, a notice, a second notice a day later, and deletion on the second day. No ticket, and nobody who has to remember.
It removes documents, messages, email threads and recorded calls, and overwrites the account. It keeps what the law says must be kept, such as signed agreements, and it logs every step. It also refuses when it should: an account with a recent paid booking cannot simply vanish.
A lot of personal data, and what stands around it
The platform holds years of conversations between students, landlords and staff. Every change to a record is written to an audit trail, a component I built and published as open source.
Consent, retention and data-subject access and deletion are built into the software, not only written in a policy. Staff sign in through the company's identity provider, not a local password. Developers never work on live data.
HTTPS from the start
Every page has been served over HTTPS from the start, with browsers told to refuse anything else (HSTS)HSTSA setting that tells browsers to use only the encrypted HTTPS version of a site for a set period. It stops an attacker on the network from quietly downgrading a visitor to plain HTTP. and old TLS versions switched off as soon as they could be.
Today visitors connect over TLS 1.3, and nothing behind that speaks less than TLS 1.2, including the link to the database.
Online viewings, days after the first lockdown
When students could no longer visit a property, I built video upload, encoding and online viewings with video conferencing into the platform. It was live days after the first UK lockdown began.
The business kept letting, and even grew slightly, in a year when most of the market fell.
A deposit lodged in two minutes, not forty
UK law gives a landlord 30 days to protect a tenant's deposit in a government-approved scheme and to hand over the prescribed information. Miss it, and the penalty can reach three times the deposit.
It used to mean thirty to forty minutes of form-filling per deposit. I connected the platform to the scheme: the data is pre-filled from what we already hold, and staff check it and confirm. Under two minutes.
A hundred integrations, and where the leads come from
Property portals, deposit schemes, payments, invoicing and accounting, credit checks, land registry, telephony, WhatsApp and SMS: more than a hundred external APIs over the years.
Portal integrations brought over 73% of qualified leads over three years. Messaging took the time to confirm a viewing from a five-day average to under twelve hours, half of them within the first hour.
Keeping it up, and moving it without anyone noticing
Uptime of 99.998%What 99.998% meansOver seven years, 99.998% uptime is about 74 minutes of downtime in total. For comparison, 99.9% would allow about 61 hours of downtime over seven years. over seven years.
In that time the platform moved from virtual machines I ran myself to containers and Kubernetes, with database replication for backups and high availability. The framework and the language underneath were replaced version by version while the site stayed open.
AI, before it was a product
Since 2014 the help centre has suggested answers while a user is still typing the question. It started as a rules-based interpreter, moved to a conversational AI service, then to workflows with retrievalRAG: retrieval-augmented generationBefore answering, the AI looks up the relevant passages in your own documents and answers from them. It reduces made-up answers and lets the source be shown.Learn more → over our own documents.
More than half of user questions are answered automatically, staff interactions fell by 53% over seven years, and far fewer leads arrive that nobody can qualify.
A team, and a second office
I opened the company's second development office, in Romania, hired the team and mentored developers at the start of their careers.
Twice a year the developers went to see the people who use what they build. Many of the best ideas for other departments came out of those days.
Taking the platform abroad
When the company looked beyond the UK, I took the platform to Singapore and Barcelona, adapted to each market.
