Some
assembly
required.

E-Commerce Experience Manager

Two sites merged, two content systems, two price lists, two sets of lens options, one customer who doesn't care about any of that. I was asked to take customer experience on after turning round the partner programme.

• First response time went from almost five working days to under 18 hours
• Unnecessary customer contacts down around 50%, through better templating and AI-assisted responses in the CMS and Shopify
• Unified returns, lens handling and service standards across both brands
• Introduced fraud prevention across both sites
• Built out-of-hours cover with a follow-the-sun model, no extra UK headcount
• Member of the senior management team

Partner Programme Manager

Rebuilt the partner programme end to end. Around 300 partners across traditional media, publishers and influencers, each needing a different approach.

• Grew the partner base 60%, aimed at the luxury brand and higher-value audiences
• Re-engaged 100 lapsed partners with targeted outreach and revised terms
• Took the programme in-house off the affiliate network, so the fees, the data and the relationships all came back to us
• Replaced flat commission with scaling payout tiers tied to partner tier and product margin. Direct costs went down. Top performers got paid more.
• Top-tier partner sales up 29% in six months
• Partner traffic revenue up 35% in three months
• Led and mentored the coordinator team, then handed the programme to them cleanly

Customer Success Leader

An early-stage SaaS startup. No customer success function, no playbooks, no operations team behind me. The product found affiliate and publisher partners for brands by mapping their competitors' partner ecosystems.

• Churn down 25% in the first year
• ARR up 75% in the first year
• Trial-to-subscription conversion up 50%
• Support tickets on key features down 15%
• Owned the full cycle: initial sale, demos, onboarding, renewals, expansions, including enterprise procurement and non-standard contract terms
• Configured every customer from scratch, starting with a session on their business rather than on our software
• Wrote the user stories, the documentation and the product videos
• When the team was made redundant and the founders brought everything in-house, I left behind playbooks for every process so they could run it without us

The best story from this job involves a global sports brand, an agency, a product that couldn't handle their regional structure, and a very large spreadsheet. Ask me in person.

E-Commerce Manager

• Amazon Marketplace sales up 75% in nine months, across a pan-European omni-channel strategy
• Merchandising lead time cut by three weeks
• Returns processing time halved

EMEA Customer Quality Manager

• Cut the diagnosis-to-fix cycle by 40%
• $2.5m in costs reduced or avoided
• Ran global cross-functional programmes with budgets around $3m
• Service delivery for enterprise, public sector and SMB customers across EMEA
• Led NPS calibration across EMEA divisions

UK/Ireland Resolutions Manager

Find out what's actually wrong first

Most "we need a new tool" conversations are really a process conversation wearing a disguise.

Deterministic where it can be, judgment where it has to be

Answerable's scoring makes no AI calls. Edition uses the expensive model for exactly one email per brand. Draw the line, then keep to it.

Keep the evidence

Every claim in the Atlas carries its source and a confidence level. Ellie's logs are append-only. If I can't show you why something says what it says, it doesn't say anything.

Check the checking

Three of the last four faults I found in the Atlas were in how I was measuring, not in the data. Now I ask what result would mean the test itself is broken.

Write it down so somebody else can run it

When Breezy made the team redundant, the founders got playbooks for every process. When I handed the partner programme over, the coordinators got a working programme, not a mess with my name on it.

Fewer steps beats better steps

Halving unnecessary customer contacts wasn't clever automation. It was noticing which questions we were causing ourselves.

Answerable

Search is turning into answering. If someone asks an AI assistant about your category, you either come back in the answer or you don't, and nobody sends you a report about it. Answerable audits how a site is represented by answer engines and scores it deterministically.

The scoring engine makes no AI calls at all, on purpose. A score you can't reproduce is a horoscope. Classification runs as a five-stage evidence pipeline: admin override, then corpus lookup, then hard rules, then weighted signal accumulation, then commit or flag for review. Sites get sorted into six types with presence modelled separately.

Built on Lovable and Supabase. Around 4,870 audits of distribution data went into rebuilding the scoring tiers.

The Atlas of Recorded Music

Music directories tell you what exists. They're very bad at telling you how anything connects. The Atlas models recorded music as a graph of claims rather than a catalogue: works, recordings, release groups and releases, with MusicBrainz IDs as the spine and every claim carrying its own source, evidence type and confidence.

The founding decision was that the non-musical links carry the weight. Rooms, hands and paperwork are first-class links, not footnotes on a release. On top of it sits Atlas Walk: start at one record and walk outwards through doors marked backwards, sideways and forwards, with your own listening history suppressing the records you already know.

Real numbers from the pilot import: Discogs credits took documented claims from 0.06% to 41.4%, added 1,496 people the graph had no record of, and Work coverage has climbed from 60.1% to 83.5% across nearly 15,000 recordings. The best-attested claims carry 46 independent sources.

The commercial shape is an open graph with a paid resolver. The claims stay public and citable. The resolution engine is the product.

FOURplay

A music discovery site built on the idea of pre-internet MTV Europe: someone with taste puts four things in front of you and you find out what you think. Daily playlists run off a seed-governed engine and lock into an archive. Feature playlists are hand-curated strands, structured as shows with three tiers.

The curation standard is the actual product. Four tracks, four different artists, discovery before rediscovery, no live versions or remasters or remixes, no decade staples, and no track picked purely because it's obscure. If you could build the list with a spreadsheet filter, it isn't an idea.

Identity and tags from MusicBrainz, genre and similarity from Last.fm, label and scene from Discogs, collaborative filtering from ListenBrainz. Directus on Supabase Postgres at the back, Astro at the front.

The Forgotten Pop Canon

A boxed set that doesn't exist yet. 55 volumes, 25 tracks each. 1,375 tracks, 1,375 artists with no repeats, 87 countries, spanning 1965 to 1999.

It runs off a written Editorial Charter, now on version 12, which exists so that every session makes the same decisions I would. One definitive track per artist, then that artist is retired permanently. Melody over rarity. "Forgotten" means invisible now, not unsuccessful then. Accuracy outranks everything.

What's left is verification, a streaming sweep, and the parts I deferred on purpose: liner notes, essays and artwork.

eSIM Sorted

Travel eSIM pricing is deliberately hard to compare. This ranks UK travel eSIM plans by actual value, aimed at the UK traveller heading abroad, with inbound visitors as a second cluster. Two connected pieces: the comparison site, and a separate ingestion pipeline built around a 57-column master schema that captures the whole UK market rather than a hand-typed sample.

Two decisions I'd defend. Genuinely uncapped "unlimited" plans rank below sized plans, because pretending to know your daily usage would be making the number up. And the paid homepage placement is labelled as paid placement, with "Top pick" reserved for whatever actually wins on value.

Ray-Ban Fan

An affiliate content site about one brand of sunglasses, done properly: 25-plus model style pages, a colourway audit framework, and the brand history behind it. Built on Grav CMS. Currently being moved to Cloudflare, which mostly means getting 12,000 product photos out of a git repository where they should never have been.

Edition

Newsletter sponsorship for independent creators with 500 to 5,000 subscribers, who have a genuinely sellable audience and no idea how to price it. Generates rate cards, media kits, sponsor target lists and pitch emails.

Next.js, Supabase, and a deliberate split on model use: cheap and fast for the structured work, the good model only for the first cold pitch to each brand. Every follow-up is a fixed template. AI where judgment is needed, templates everywhere else.

Ghost Core / LE-914 "Ellie"

A Toymax RAD toy robot from the 1990s, being retrofitted into something that actually listens, remembers and answers back. Local models via Ollama, vector memory in ChromaDB, Whisper for speech, HomeKit so it can be usefully annoying about the lights. A Raspberry Pi Zero 2W runs the body inside the original chassis and talks to the brain over WebSocket.

There's a personality constitution, because a droid with no opinions is just a speaker. Archivist and veteran, with precision sarcasm. The core build rule: logs are append-only, nothing is ever deleted. A droid must be able to testify.

Race Night

A race night app for a bar in Edinburgh. Host screen on the big telly, punters betting from their phones, real cash tote with stakes tracked and payouts calculated. Configurable fields of four to twelve horses, a "buy a horse" mechanic, and a traditional bookie's board for the look of the thing. Record Roulette, the version where the runners are records and the race is a listening game, is parked behind it.

The Forge

The thing that stops all of the above from quietly dying. A ranked portfolio board that scores every project across four dimensions with a neglect penalty, so the project I've been avoiding climbs on its own. It rebuilds itself every evening and tells me the single most urgent thing. My training programme sits at the top of the same page, on the grounds that it's the same problem.