Never tell a developer they're doing great: Lessons from a Data & AI Leadership Conference
On scaling engineering teams; the sexiest new job in AI; and why data is LEGO... or is it... pizza?
Iām a nerd. In German youād call me āStreberā. When I go to a conference, Iām that girl in the front row tapping away furiously on my laptop, afraid to miss a thing.
Afterwards Iā¦ I ā¦ well, I never look at my notes again. Until now. Today, in a break from my usual posting style, I want to share my key takeaways from the Data & AI Leadership Day, which my company (Xebia Data) held in ZĆ¼rich this week. Iāve grouped them per talk, so feel free to skim over the headings and sub-headings to find the lessons that are useful to you. Enjoy!
Marketplace-driven Innovation in Data as a Service (DaaS) Businesses
A better title would have been āhow to make tonnes of money from your data.ā In the blurb to his talk, Snowflakeās Fawad A. Qureshi promised free pizza if we didnāt like it. Luckily, I wasnāt hungry. Hereās what he had to say:
Data products are like LEGO bricks. Or, more accurately, theyāre like Rebrickable, a website where you can find patterns for re-using your existing LEGO bricks for new creations. What Fawad means is, there is value to be found in taking something from one industry and applying it to another. For example, an aviation company might try to create a data product out of their tourism data.
No, wait, data is like a pizza. Data can take multiple forms, depending on how ready-to-consume it is: Pizza dough is the raw data you serve for others to use as they please; Frozen pizza is the data they just need to heat, like your cleaned, transformed, documented Big Query database; And a ready to eat slice is your easy-access data product, like a self-serve dashboard. Businesses should be aiming to build something consumable, but the problem is, 90% of data out there is garbage: system logs saying āIām hereā, for example. In Fawadās words, āforget pizza dough, this is the raw flour grains!ā How many people can take raw pizza dough and convert it into something good? Not many. But smart data businesses can do it. For example, those logs could be turned into traffic or usage data, and that can be useful.
You are wasting the data you have. Fawad is convinced that companies arenāt using enough metadata in their internal processes. For example, he could ask his clients "what is your most profitable data feed?" and they have no idea: they email out a CSV and then have no idea what happens with it, or what potentially-valuable use-cases the recipients have discovered. This is a huge, wasted opportunity. And similarly, Fawad can ask "what's the data stream you can drop without customers starting to scream at you?" and his clients don't know.Ā Thatās a huge problem, if it cost resources to create that data stream in the first place.
The value of data lies in having just one copy. Moving on from LEGO and pizza, remember when we used to rent VCRs? (Iām just old enough!) This was an ETL tool. The problem was, it wasnāt scalable. There was a limited number of any given video at any one time, and when they were all rented out, that was too bad. Later we had CDs, and we generated value for ourselves by burning them for all our friends, at a loss to the original data provider. Not good. Then Netflix came along and said, "we'll make a single copy of the data, you bring your own compute and bandwidth ā and pay us for the privilege!ā Genius, right? And what made it possible was that there was only one copy of the data, but it could be served to multiple, paying customers, at once. So what Fawad is actually saying is, you donāt just want to build a data product that serves one customer at a time. You want to build a platform. Allow me to explainā¦
Aim for multi-sided data platforms. According to Fawad, everything is a platform these days ā or at least, it tries to be. "Just because my tech has two capabilities, it's a platform," he says, mockingly. No. It needs to have a network effect, a two sided market. What's that? Uber, for example: it has the rider and driver app. The more one increases, the other increases. The moment wait time goes up, you switch apps. Food delivery services are similar, but they're triple-sided: customers, restaurants, markets. The demand and supply of any one effects that of the others. You want to build a data platform where demand and supply spirals perpetually upwards.
The future is full of data-centric apps, and you should be one of them. Snowflake are heavily pushing the concept of a ādata marketplaceā. In their words, āa data marketplace (or data sharing marketplace) is an online, transactional store that facilitates the buying and selling of data.ā As a business that owns some data, you want to become a marketplace provider.Ā And the reasons are perpetual, going beyond the revenue to be gained from selling the data you already have: you can use the usage metadata of the data you sell to find more opportunities to create value. That is, internally, you might only have a few data scientists. But once it's in the marketplace, then by learning how consumers are using the data, you can copy that idea to make (and sell!) a new data product.
So how can you find data sources that would suit a marketplace?ā¦ Think about the organisational culture, the entire value chain, how you do assessment, monitoring, and so on. Analyse the data flows from every angle. Put yourself in the shoes of different types of consumers. This is where youāll strike the real gold dough. Pun absolutely intended.
You need to make your data discoverable by Large Language Models: Fawad predicts that LLMs will soon help guide people's browsing on data marketplaces. Thus, having informative metadata for the LLMs to access will be vital. And even within your own company, LLMs can help you move towards becoming a Data as a Service company. For example, if you've built a data governance process and have good metadata, you can expose it to an LLM service, literally asking it a question like āI'm looking for a last-mile optimisation service for my delivery trucks, what information could be useful for such a service?ā The LLM might well be able to discover semantics in your metadata you never even thought about, and reveal exciting new data use-cases and potential data products to you. Insider tip: Fawad hinted that Snowflake is going to do something along these lines within the marketplace within a few weeks, so Iād be watching this space. Update: Sure enough, Snowflake just announced its new Document AI, allowing users to query data using natural language, at its 2023 summit.
AnalyticsĀ Translator: From idea to impact
I loved this talk so much from UBS Bankās Dr. Michel Neuhaus, I pretty much photographed every slide. Allow me to brain-dump on you:
Move over, Data Scientists, this is the new sexy. The analytics translator is a dedicated role that sits in the space between business leadership, IT architecture and data science. The purpose is to ensure business alignment and delivery success, and according to every consulting company under the sun, you need to have one.1
What does this look like, in practice? Analytics translators write business cases, act as intermediaries between product and development teams, discuss roadmaps and use cases, challenge ideas, and more. In smaller companies or teams, one or more people might step into the role if and when itās required. UBS, a much larger company, has a team of permanent translators. But even they donāt have the capacity to work on every project; thus, each translator usually focuses on 1-2 large projects, where success is critical, and steps in on 1-2 smaller ones when their advice is needed.
Be obsessed with relevance, feasibility and impact potential. If you struggle with any of those points, you probably need an analytics translator, to help you iterate on proposed product ideas until you get there. (As a vital side-note, remember that impact only matters if you can bring it live).
What does it take to become an Analytics Translator? How do you upskill employees into this role? Dr Neuhaus confessed that it's hard to find good people. An analytics translator needs to be technical and have domain expertise, otherwise theyāll never be able to convince their coworkers on both sides. You could use a senior data scientist who likes to tell stories, or a business person who wants to get into data topics. Various online courses exist, and there will no doubt be more, as the importance of this role is realised. If youāre in ZĆ¼rich, by small chance, Dr Neuhaus recommended the Certificate of Advanced Studies in AI Management, from my esteemed Xebia colleague, Afke Schouten.
Ask the right questions, before you jump into Generative AI. As was inevitable, Dr Neuhaus faced an audience question about the impact of Generative AI in finance. His opinion? It's going to change hundreds of problems, from minor to major.2 However, Dr Neuhaus stressed the importance of challenging proposed Generative AI use-cases and projects. Exactly what the right questions will be depends on role and organisational unit, such as Business Leader versus IT Architect versus Data Scientist. For example, a Data Scientist might have heard that Generative AI works much better than BERT, and be keen to try it. Before doing so, she should ask: Does it solve a real problem? Do we need explainability? How will we bring it live?
Is the Use-Case Worth It? A Rubric. Speaking of asking the right questions, Dr Neuhaus shared his system for evaluating the value of a proposed AI use-case. It focuses on four areas and some key considerations within them. Iāll include the slide here, but leave it to your entrepreneurial and data-driven minds to think about how these dot points could be relevant for your company:
Empowering Engineering Teams for Scale (Keynote)
Roche is a global, Swiss-based company specialising in pharmaceuticals and medical diagnostics. Deepak Sondur leads their cloud and edge platform strategy, and used his talk to describe critical factors to developing and scaling digital health products. Here are some key points I (frantically) noted down:
Donāt tell your engineers that theyāre doing a great job. Or, rather, just donāt tell them all the time. Why not? It will stop them feeling challenged, which can have one of two disastrous outcomes: on the one hand, people might get comfortable, theyāll lower their standards, and productivity will drop; on the other hand, theyāll become demotivated, seek inspiration from outside the company, and ultimately, leave. In fact, Iāve just started āAmp it UP: Leading for Hypergrowth by Raising Expectations, Increasing Urgency, and Elevating Intensityā, by Snowflake CEO Frank Slootman, and I have a feeling his advice is leading in the same direction.
You need a Security Champion. someone who acts as a bridge between developers, QA and security teams, in order to help share knowledge and disrupt silos and bottlenecks. Generally, this will be an engineer who evangelises and teaches security best practices among the development team, and translates security concerns into terms developers can understand and work with. For advice on setting up a security champion program in your organisation, see the Open Worldwide Application Security Project (OWASP)ās excellent guide.
A shared responsibility principle within engineering teams is essential, especially during crises, such as security alerts. This way, you avoid the blame game and can concentrate on getting all hands on deck to resolve the issue.
If you want to be really successful in creating a product, focus on developer experience. This means four things: developers should put themselves in the shoes of their consumers (e.g. the users of the BI dashboard they built). They need to automate relentlessly. And they need to self-service everything, and open up their toolchain for other developers. Combining these key elements will help software solutions be more broadly adopted, which will increase demand for the solutions to be adapted and scaled. At the same time, developers will already have the tools and processes in place to make that scaling a whole lot easier and more successful.
Bold and Responsible in the New Era of AI
This talk by Googleās Vladimir Vuskovic turned out to be more of a showcase of the incredible advancements the company has made in itās Generative AI capabilities in recent months. Iām already working a lot with their Generative AI Studio, so I didnāt end up taking any notes. But I did jot down Vladimirās recommended watching, when asked for his favourite resources on the state of AI and AI regulation: āState of GPTā, from Microsoftās developer YouTube channel. I can also leave you with a Google-AI-generated joke about mountains, which did actually make me chuckle:
What did the angry parent mountain say to the child mountain?
Don't you give me that altitude!
Hilarious.
HBR on McKinsey.com: Analytics translator: The new must-have role