Chris Walker Data and analytics leader Exeter, UK

I find out what's actually going on.

Then I help the business see it, understand it and act on it. Sometimes that starts with a blank page. Sometimes it means picking up something half-built and working out what it needs next. I'm at home with either.

Most questions reach a data team already tangled: a product decision wrapped round a pricing problem, wrapped round a number nobody quite trusts. Pulling harder doesn't help. I work out which thread to start with, what to save for later and what to leave alone, and I bring the people who own the decision along with me.

For the last five years that has been at Overleaf and Digital Science, as Head of Analytics and then Director of Product Analytics. I built the data platform and the team, and I still write the SQL. Before that I was a systems engineer, ran two businesses of my own, and spent six years as the marketer asking for the numbers. I've sat on both sides of the table.

Now
Director of Product Analytics, Digital Science. Six products, five market segments, nine-figure revenues.
Before
Head of Analytics, Overleaf. Built the function, the platform and a team of six.
Most use
Where there's a real business and a real product, and more questions than trusted answers. Mature data foundations optional.
Works with
Product, commercial and leadership teams. Board to engineers, in the language each of them uses.
Hands-on
BigQuery, Dataform (dbt), SQL, Python, Looker, Mixpanel, and agents running in production.
Education
MEng, University of Cambridge.

How I go about it

Eight habits I keep coming back to. Each one shows up in the work further down the page.

Find the real question.

The one that arrives is rarely the one that matters. Sales asked for more leads, and the leads were already customers. A test said "win", and the plan it promoted was losing money.

Start with the end in mind.

Before jumping in, work out where we're going and what success looks like. I learned it as a systems engineer, where getting paid meant delivering something that would still conform years later. The thinking goes at the front.

Focus first.

Quality over quantity. At Overleaf the business had 25+ questions. I chose five and left the other 80% for later. Cut the scope if you have to. Don't cut the standard.

Think once.

Don't solve the same problem from the beginning every time. Solve it properly, capture the thinking as a definition or a model, and reuse it. It's what makes self-serve possible, and agents too.

Trust matters.

The right answer isn't enough if people don't believe it. I'll go a little slower to be robust, and hold that in tension with the need to move. "Can we trust this?" gets answered first.

Equip the decision.

My job is to give decision-makers everything they need, facts first and opinion last. Sometimes the right call is to change nothing. It should still be a call someone made on purpose.

If nobody uses it, it didn't happen.

I judge the work by what people do with it: a dashboard tab that stays open, a metric a squad plans around, a checklist the team follows without me.

People keep the judgement.

Agents and automation take the routine. Deciding what to measure, working out why it moved and saying what to do about it stay with a person.

Five times it wasn't what it looked like

Each of these started as one problem and turned out to be another. The last line of each is what's true now.

2021 to 2024OverleafPlatform and experimentation

From 25 questions to 1.5 billion events a month

It looked like
Overleaf needed analytics. Everyone had questions, nobody had answers, and the obvious move was to buy a tool and start building dashboards.
What was going on
The data lived in silos: three backend databases, a payment provider and Salesforce, with no way to connect what users did to what they paid. And there were 25+ questions on the table, with no agreement about which mattered most.
What we did
I gathered the questions from across the business, grouped them into themes and chose the first five to answer. That was enough to prove the idea and show what the platform had to do. It went up in layers on BigQuery and Dataform: raw data as received, a clean layer, then a modelled layer the business could use. A split-testing method went on top, so product decisions could be tested and not argued.
Now1.5 billion events a month, a team of six, and 94 experiments run through a framework that became part of how product ships. Every other story on this page runs on it.
SourcesThree layersUsed by Backend (3 databases) Payment provider Salesforce Exchange rates Sister products Rawexactly as received Cleanstandardised, Dataform Modelledbusiness-ready Product squads Sales, marketing Finance Exec and board 1.5B+ events / month
BigQuery underneath, Dataform (dbt) for the modelling, Looker and Mixpanel on top.
2021 to 2024OverleafWays of working

A seat at the table

It looked like
A new analytics team, and a product organisation ready to send it requests. The usual model: a queue, a dashboard, a thank you.
What was going on
By the time a request reaches a queue, the question has already been framed, often around the wrong thing. An analyst is most useful in the room where the problem gets defined.
What we did
I positioned analytics as the quantitative mirror of design. Design brings the user's voice. Analytics brings what users actually did. Analysts became mandatory in squad planning every Shape Up cycle, sometimes as a fourth member of the product trio, sometimes in place of a designer or an engineer.
NowOn a project about compile errors, the analyst in the first planning session redirected the work from the interface to the compile logs, and saved a good deal of user research. The model spread to more squads, and analysts have been part of planning ever since.
TrioQuartet Prod. Design Eng. analytics: reports afterwards Dashboard over the fence Prod. Design Eng. Anlst analytics: in planning Six-week Shape Up cycle Design brings the user's voice. Analytics brings what users did.
The analyst mirrors the designer: present when the question is framed, not after.
2022 to 2023OverleafPricing

The winning test that was losing money

It looked like
A success. A change to the upgrade page had been A/B tested, and subscription starts and revenue per user both went up.
What was going on
Over the following months people were choosing the cheapest plan in place of the ones above it, and losing the collaboration features we most wanted them to use. That plan grew to over 7% of billings. Lifetime-value analysis put the cost at five figures a month, and compounding.
What we did
My job was to equip the people making the decision. I set out four options with the impact of each and took the CEO, CFO, CMO and CTO through the analysis, facts first, answering "can we trust this?" before anything else. Keeping the plan would have been a reasonable call if the strategy had pointed that way. It didn't.
NowThe plan was retired. ARPU and lifetime value moved as the model predicted, and subscriptions fell less than anyone feared. Tests aren't designed to catch an effect like this, and we still need to test quickly. So where a change carries a risk of one, we now follow the longer-term trend after the test ends.
What the test saidWhat the model said Short-term conversion up looks like a win Cannibalising higher plans 5-figure monthly cost Four options, each costed Retire it Harder to reach Upgrade gateway Double down Taken to each of them, facts first CEO CFO CMO CTO Their decision: retire the plan the forecast held
Four options, four executives, one decision. Theirs to make.
2023 to 2025OverleafSales enablement

The customers sales couldn't see

It looked like
A thin B2B pipeline. Sales needed more leads.
What was going on
The leads were already customers. Thousands of people inside universities and companies were paying individually, and the evidence was spread across four systems nobody could join. Preparing for one call took hours, and often an engineer.
What we did
Discovery first, then an MVP with the five numbers the sales team actually used. The awkward transaction types were left out and labelled as a limitation. It worked, and people wanted more. I didn't try to do everything and do it badly. I built the business case and held the tool steady until the business was ready to back it with a dedicated analyst. Then we rebuilt it properly.
NowAround 50 people use it every month, and its reach goes well beyond the sales team it was built for. It surfaced large organisations nobody knew were customers, and the sales team say they "always have the tab open".
Sources One model Adoption Payment system Analytics events Production database User affiliations B2B sales dashboard 5 metrics first 50 monthly users beyond sales
An MVP first, then a business case, then production. Eighteen months end to end.
2026Digital ScienceAgentic systems

Agents for the routine, analysts for the judgement

It looked like
A capacity problem. A backlog of data modelling nobody had time for, and a steady stream of questions from product teams.
What was going on
Most of those questions were routine, and the routine was crowding out the work only an analyst can do: deciding what to measure, working out why something moved, and telling a product team what to do about it.
What we did
It started as two experiments against live BigQuery: an agent that answered questions in plain English, and one that drafted metric designs from GitHub issues. What made them worth trusting was the groundwork underneath: a well-modelled metrics layer with agreed definitions, so the agent reads from the same answers an analyst would. Then came a framework for the harder cases, so that when the answer isn't sitting in the metrics layer the agent knows how to find its way round the warehouse and dig.
NowPeople ask a question in Slack and an agent answers it, from the metrics layer where it can and by digging deeper where it can't. Drafting a design from an issue, or building the model from one, is routine. Most of the product teams' questions are handled this way, by agents the team built and maintains, and the analysts spend their time on more impactful analysis.
Questions Question in Slack Agent Metrics layeragreed definitions if it isn't there Warehousea framework for digging deeper answer, back in Slack Modelling GitHub issue Agent drafts Analyst reviews Model built a person decides
The metrics layer is the part that makes the answers worth trusting.

The numbers

At work I sort metrics into four tiers: north star, KPI, health and diagnostic. Here are the same tiers, applied to me. Every figure is one I can take you through in detail.

North star
50

People using the B2B sales dashboard each month, well beyond the team it was built for. Adoption is the measure I care about most.

North star
94

Experiments run through the split-testing framework that became part of how product ships.

KPI
1.5B+

Product events a month on the BigQuery and Dataform platform I designed.

KPI
0 to 6

The analytics team at Overleaf, from me alone to six across data engineering and analysis.

Health
40% YoY

Revenue growth at Overleaf over the same period, with 25% user growth alongside it.

Diagnostic
8h to <1h

Turnaround on flight-trial data analysis after I rebuilt the tooling, back in 2010. The habit started early.

Four jobs before this one

I didn't start in data. Each of these left me with something I use every week.

2007 to 2015 · Honeywell, Selex ES, J+S

Systems engineer

Defence and aerospace teach you to get things in the right order, and to respect a system where a wrong answer is expensive. On a flight-trials programme I rebuilt the analysis tooling and took turnaround from eight hours to under one, so the team could recalibrate and fly again the same day. It won the company's innovation award.

2014 to 2021 · Potent Products, Potent Digital

Founder

Two businesses with my own money in them. One imported and sold a children's moneybox under exclusive UK and European rights, and reached six-figure turnover as a part-time venture. You read a P&L differently once it has been yours.

2015 to 2021 · Potent Digital

Marketer

Six years as the person asking for the numbers: acquisition, SEO, conversion and email for startups and small businesses, with six-figure Google Ads budgets. For one publisher, asking readers what they wanted to hear about lifted open rates six to seven times and took click-through past 60%.

2020 to 2021 · Overleaf

Product manager

I came into analytics from product. I know what a squad needs from a number, how late is too late to bring it, and what it's like to make the call without one.

20052010201520202025
University of CambridgeMEng, 2003 to 2007
HoneywellProject Manager, 2007 to 2008
Selex ESSystems Engineer, then Senior, 2008 to 2013
J+S LtdSenior Systems Engineer, 2013 to 2015
Potent ProductsFounder, 2014 to 2019
Potent DigitalFounder and Digital Strategy Consultant, 2015 to 2021
City Life Church ExeterChair of Trustees, 2016 to 2024
OverleafProduct Manager, 2020 to 2021
OverleafHead of Analytics, 2021 to 2024
Digital ScienceDirector of Product Analytics, 2025 to now

Outside work, I chaired the trustees of City Life Church Exeter for eight years, from the charity's foundation.

If something's tangled, I'm happy to take a look.

Open to conversations. I've worked remotely from Exeter for the last ten years.

Email chriswalker85@gmail.com
LinkedIn linkedin.com/in/chris-walker-uk
Based Exeter, UK