From Data Analyst to Senior DS Manager at Skyscanner
How an engineer wandered into data, grew into a modeller and ended up running teams at Skyscanner.
Fellow Data Tinkerers,
Following on from previous posts talking to people in the field, today I will be talking with
who is a Senior Data Science Manager at Skyscanner and writer of the Senior Data Science Lead newsletter.
We talked about his rise from data analyst to Senior DS Manager at Skyscanner, what “production-ready” really means and why the real intelligence in data science lives before and after the model.
So without further ado, let’s get into it!
Can you tell us about your role?
I’m a Senior Data Science Manager at Skyscanner, the travel marketplace that helps millions compare flights, hotels and car hire around the world. I currently oversee three data science squads though I don’t manage them directly day to day. My focus is on tribe health: making sure the broader data science org runs smoothly through smart hiring, resourcing and clear ways of working.
I also spend a lot of time connecting dots across teams by translating between strategy and delivery with directors and VPs. These days, my role leans more toward systems and strategy than hands-on modelling which keeps things interesting.
If you’re unsure how data science management differs from Individual Contributor (IC) work, check out my article below.
How did you break into data science?
I got into ‘data science’ before it was cool 😎 I studied mechanical engineering then a master’s in robotics which was my first deep dive into maths and code. That’s where I fell in love with the mix of logic and creativity data work offers.
My first role was as a data analyst at Deloitte: lots of Excel but little experimentation. So I started self-learning machine learning on the side: Andrew Ng’s course, textbooks and endless projects in R and Python. That eventually landed me a junior data scientist role at an insurance company back when models meant GLMs and careful feature engineering, not plug-and-play XGBoosts.
From there, I worked my way up: junior, senior then manager. Looking back, I broke in through self-study and stubborn curiosity. It was a patchwork education built from late nights and trial and error. Today’s structured bootcamps and degrees would have made it much easier.
Jose’s path
data analyst → junior data scientist → senior data scientist → data science manager → senior data science manager
How has your perspective changed since moving from contributor into management?
The shift from IC to manager completely rewires how you see data science. As an IC, progress is all about building whether coding, experimenting or delivering something tangible. As a manager, you move one step back: you enable others to build but stay technical enough to guide, challenge and represent your team effectively.
I often think about the shift across three dimensions:
From doing to enabling. You stop solving problems yourself and start creating the environment where others can by allocating resources, sequencing projects and removing blockers.
From depth to breadth. You move from being deep in one model to connecting work across teams and defining what ‘done’ actually means.
From output to people. This is the biggest change. Your real system is not the pipeline anymore, it’s the team. So your job is to focus on alignment, motivation and career paths of your team members.
You still need technical depth, but success shifts from what you build to what your team sustains without you.
If you are thinking of transitioning from IC to a DS manager role, I have shared my thoughts in the article below.
What does a typical week look like for you?
My week now revolves around conversations and building relationship rather than focusing on code. So to stay sane, I try to give each day a different flavor.
Mondays are for people. It is a day packed with one-to-ones focused on growth, feedback and general wellbeing. It is a brutal day in terms of energy but I know what I am signing up for - it sets the tone for the rest of the week.
Tuesdays and Wednesdays are project days. That means discussions with my managers, stakeholder reviews and executive updates. Less about personal growth, more about delivery, progress and course correction.
Thursdays are my office days, which I really value. Even if I am still in meetings, the quality of conversation is different. You notice things — tone, energy and side comments that never come through on Zoom.
Fridays are leadership and thinking time. It is when the data science leadership group meets, and when I try to carve out mental space to think strategically about team structure, hiring or what our next quarter should look like.
So yes, it’s meeting-heavy but every conversation is a piece of the system I’m building.
What kinds of data do you work with?
Almost everything we do supports one goal: helping travellers make better decisions. That means three broad types of data:
Structured data - the classic kind: rows, columns and numbers. Think flight prices, hotel ratings, click behaviour and search patterns. We use this to power ranking algorithms, optimise search and predict trends like when prices might rise or fall.
Unstructured data - text that lives outside tidy tables. For example, descriptions of hotels, destinations or reviews. We ingest and structure that information so that our models can learn from it by turning travel text into something measurable.
Image data - photos of destinations and hotels. We work on tagging and analysing those images to understand aesthetics, identify objects and even predict which photos help users make better choices.
If it helps travellers pick the right place or time, it likely ends up in our pipelines.
How does this data tie back to business outcomes?
Skyscanner is a marketplace, so data has to serve two sides: users and partners. Most of our data work exists to strengthen that balance.
For users, we make search faster, cheaper and more relevant by surfacing better deals and clearer choices which builds trust.
For partners, we share anonymised insights about traveller behavior to help optimise pricing and availability.
In short, data is our nervous system. It keeps both sides informed, aligned and confident in the marketplace.
What’s your process or checklist before calling a model “production-ready”?
For me, A model is production-ready when you could disappear for six months and still know how it works, how it performs and how to fix it without needing the person who built it.
At Skyscanner, that readiness happens in three broad phases:
Exploration. We start with experiments. proofs of concept or A/B tests that show the new system actually outperforms the old one. During this phase, notebooks are fair game. They are perfect for ambiguity and rapid exploration. But as soon as direction becomes clear, we move to proper repositories and code structure. The goal is to avoid that painful rewrite later, the one where a '“clever” Jupyter notebook has to be refactored into a real pipeline.
Hardening. Once a model proves its worth, we secure it. That means code reviews, consistent repo structures and clear metadata. We use MLflow in Databricks for tracking and versioning, tagging models from staging to production and recording lineage - which data it uses, where outputs go and how it performs. It is about wrapping the model in enough context so it can live independently.
Monitoring. Finally, we integrate with our MLOps monitoring services, define metrics for drift and performance and set up alerts. The idea is that every model we ship should be observable and explainable.
This is our version of going from bronze to gold. In short, production-ready means the model should work today and it can survive tomorrow without surprises.
What’s in your tech stack?
Our stack is fairly modern and pragmatic - nothing exotic, just what works well at scale.
Languages: Mainly Python
Data and compute: Databricks on Spark (mostly PySpark)
Experimentation: In-house Bayesian platform
MLOps and model management: MLflow and Cortex for tracking, deployment and lineage
There are a few specialised vendor integrations here and there but these are the backbone of our daily work. It’s a setup that lets us move quickly while keeping the discipline needed for production-grade machine learning.
How do you use AI in your day-to-day work?
I would say my use of AI these days is fairly pragmatic. I do not code much anymore so I am not the classic “pair-programming with ChatGPT” example. For technical reviews or performance checks, I still prefer doing things myself.
Where AI does help me is in three main ways.
Writing and communication. I enjoy writing — it is how I think — so I do most of it myself. But when something needs to land well with stakeholders, or I need to rephrase a tough message, AI is a brilliant second set of eyes. It helps me refine tone or framing without losing my voice.
Research and exploration. This is where it truly shines. I often use AI to surface what’s out there — relevant papers, blogs, or case studies around a problem we are facing. Before, that meant endless Googling and chance discovery. Now, I can get a solid overview of the “art of the possible” in minutes.
Challenging ideas. I sometimes describe an ongoing project to an AI model and ask it to stress-test the logic — “What angles am I missing?” It does not know our internal context but it forces new ways of framing the problem, which is often what leadership work really is.
What I do not use AI for is note-taking or meeting summaries. I tried that and hated it. Writing my own notes helps me think and remember and no tool can replace that.
How do you manage multiple senior stakeholders?
For me, stakeholder management happens on two levels: projects and relationships.
The project side is the easy bit. If you make updates part of your process and over-index on communication - progress, blockers, delivery status - you buy yourself credibility. That regular rhythm makes the harder part, the relationship work, much smoother later on.
Relationships are where the real leverage sits. Senior stakeholders care deeply but they are chronically busy. They rarely have the time to chase context so you have to bring it to them. I make a point of doing monthly one-to-ones with relevant directors and VPs across product and marketing. They are only half an hour or an hour but they create trust, shared history and a sense of causality - the “why” behind every update.
The trick is to make it effortless for them to engage: come prepared, bring context and make every conversation valuable. That’s how updates turn into alignment.
Senior stakeholders care deeply but they are chronically busy. They rarely have the time to chase context so you have to bring it to them.
What’s a common misunderstanding about data science?
From the stakeholder side, there is still a common belief that data science equals machine learning - that our job is to build the clever model that spits out predictions. It is a flattering myth but a wrong one.
Building models today is the easy part. AutoML, ChatGPT, and a dozen other tools can already scaffold a decent model. The real skill is problem translation - understanding the business context deeply enough to turn it into something useful. Sometimes that solution is an ML model but sometimes it is a heuristic, a rule system or even a better metric definition.
That is why data science teams should be involved early, before a problem is framed as “we need a model.” We are partners in defining success, testing feasibility, and shaping what ‘good’ even means.
From the data scientist’s side, there is another trap: falling in love with the model itself. The craft of modelling is becoming commoditised. What sets strong scientists apart now is their grip on everything around the model - feature design that captures domain reality, validation that exposes blind spots and the interpretation that links outputs back to decisions.
The middle bit - the model - is just plumbing. The intelligence lives before and after it.
The real skill is problem translation - understanding the business context deeply enough to turn it into something useful.
One thing you wish you had known earlier about data science
Early in your career, it’s easy to chase technical perfection - the perfectly tuned model, the spotless codebase, the elegant solution. I did it too. And I learned that the perfect model is useless if it never reaches production.
My first piece of advice: do not seek perfection - seek impact. A simpler model that goes live and adds value is infinitely more useful than a masterpiece that stays in a notebook. Ship early, learn fast and focus on understanding the end-to-end path to production. That experience teaches you far more than another 0.1% accuracy ever will.
My second piece of advice: over-index on communication. Data science does not exist in isolation. Keep people informed - stakeholders, engineers, PMs. Share context, flag bottlenecks and make sure no one is surprised by your progress or your problems. “No surprises” is a good north star for collaboration.
If I’d learned those earlier, I would’ve saved a lot of frustration.
A simpler model that goes live and adds value is infinitely more useful than a masterpiece that stays in a notebook.
Any spicy takes?
Yes - and it is about AI itself.
I think we are living through one of the greatest technological shifts since the internet. The rise of LLMs has democratised capability in a way that feels almost supernatural. But at the same time, I suspect this era will become one of the great overestimations of our lifetime - not because the tech is bad, but because the story around it is.
The industry’s loudest voices keep promising AGI as if scaling the next model by another trillion parameters will somehow summon consciousness. But physics and economics do not bend to marketing decks. Each new model gets exponentially hungrier while delivering diminishing returns - we are burning data centres to gain a few decimal points of improvement.
Even if we had infinite GPUs, we would still be training on words, not the world. These systems have no concept of causality or physical intuition. They do not “understand” gravity; they merely know that “apple” often co-occurs with “falls.” They are brilliant at pattern completion, not truth discovery.
The real revolution is not the mythical AGI, but the democratisation of intelligence - the fact that anyone, anywhere can now wield reasoning tools that used to belong only to experts. That changes everything about work, creativity and access. But it is augmentation, not awakening.
So yes — LLMs will reshape our century. But the singularity? Probably not. We are not creating gods; we are mass-producing very eloquent parrots. And honestly, that is still pretty extraordinary.
If you enjoyed reading this, check out Jose’s newsletter for his 5 part series on LLMs and how they are shaping our work.
Was there a question that you would like to ask?
Let me know your thoughts by replying to the email or leaving a comment below!
If you are already subscribed and enjoyed the article, please give it a like and/or share it others, really appreciate it 🙏
Keep reading
From Dental Cleaning to Data Cleaning: How I Pivoted to Healthcare Analytics
Meet Thais Cooke who is a Senior Healthcare Data Analyst and writer of the Journal of a Data Analyst newsletter.
We talked about her unplanned pivot into data, what healthcare data analysis look like and how she thought she was being scammed by a LinkedIn ‘Impostor’.
From Marketing to Data Engineering: How I Made the Switch
Kicking off a new series where data folks share how they got here, what they’ve learned and how they actually work day to day.
First up: Alejandro Aboy, Senior Data Engineer at Workpath and writer of The Pipe and The Line
We talked about his path from marketing to data engineering, why his teammates call him an octopus and his take that “big data” is a myth for most teams.
 | A guest post by
|