Interview with Diogo Oliviera
Machine Learning Team Lead at Loka
Description of Current Work
​
Q: Can you describe your position at Loka AI and what your normal day looks like?
​
A: I’m a machine learning manager. My typical day is split between engineering tasks and managing tasks.
On the managing side (the least interesting, in my opinion): organizing the team—we have almost 80 ML engineers—and assigning each engineer to a project, ensuring the right people are on the right projects. I review all SOWs we sign, making sure whoever wrote the SOW followed normal procedure and set reasonable timelines. I also handle salary raises with my team and similar topics.
​
On the engineering side: given the current AI boom, we do a lot of generative AI work—everything from chatbots, intelligent document processing and ingestion, to whatever else involves GenAI. We have projects with studios that want to generate parts of images for advertisements rather than shooting everything. We’ve even worked with hairdressers who want to preview how certain hairstyles look.
​
Then there’s the more “traditional” ML/DL: classification with tabular data, computer vision, and so on. At Loka AI we try not to work alone—there are always two people on a project. I’m paired with someone, and we’re coding: either pushing products toward production or writing a small POC to prove the tech can handle something more complex. If we’re not sure it’ll work, we code it to demonstrate to the customer so they can show their users and decide if it brings real value or is just expensive compute on top of their platform.
​
Company Overview
​
Q: Can you talk a little about Loka AI—what the company focuses on and its mission?
​
A: I really like our vision: “ideas that ship.” People come to us with ideas—some clear, some random—and we bring on engineers to help ship them. When I started, we mostly worked with startups. We’ve built entire startup products in-house, then helped them staff and take over so they own the product.
We’re essentially a consultancy: we build useful things and hand them over. We don’t want to be the long-term maintainers; we want customers to truly own what we delivered.
​
Career Journey
​
Q: Can you give a quick overview of your career journey—education and how you got into the work you do now?
A: I studied electrical and computer engineering in Portugal. I went in thinking I’d do bionic arms and processors—for no real reason. In school I did electronics, programming, math, physics, the usual.
​
At the end of my bachelor’s or the start of my master’s, I interviewed for a summer internship. I told the engineer I liked solving riddles but hated repetitive work. He said, “You’ll like machine learning—here’s a course.” I tried Andrew Ng’s ML course on Coursera, loved it, and woke up an hour early every day that summer to do it. From then on I was set on ML.
​
On Erasmus in Denmark, ML lived under computer science, so I switched to take ML courses. I did my master’s thesis on human pose estimation. After graduating, some companies said, “We have ML, but we can’t promise you’ll work on it.” I could say no. Then a startup looking for pose estimation expertise reached out via my advisor; I consulted part-time on my thesis topic.
​
Loka contacted me for a backend role. I told them I wanted ML, not backend. They insisted I interview; I said I’d show up and tell them I wasn’t interested. In the interview I asked if they had anything in ML; they were creating the ML team. I joined as an intern, then a junior, and I’ve been here five years. We grew from 4 people to ~80—things keep breaking, and we keep fixing and improving. It’s been fun.
​
Future Opportunities
​
Q: What are the most interesting ML career opportunities emerging over the next 5–10 years?
​
A: Hard to predict—things change year to year. Take this with a truckload of salt. One of the most important skills is critical thinking and evaluation: designing good experiments, using a scientific method to test whether something is true, and understanding how to evaluate models.
​
GenAI is uncertain; rules aren’t fully set. A high school student can build amazing things if they code a bit—like Legos. The hard part is ensuring what you build doesn’t break in slightly different scenarios. Creating the systems around your model—guardrails, tests, evaluation suites—to check correctness is where you add a lot of value. Tech stacks change; you can relearn them. Solid math helps, but the essential, hard part is thinking logically about reliability, failure cases, and when not to use a tool.
​
Advice for Students
Q: Any specific advice for someone trying to work in ML?
​
A: What I’d do differently: I focused heavily on theory and grades. Theory matters, but I wish I’d learned proper coding practices earlier—clean structure saves pain later. College did teach me how to get out of tough situations: someone throws a big problem at you, and you figure it out.
​
My suggestion: build things—any random project. Start simple. You’ll quickly see real problems and learn faster. Theory explains why, but building teaches how. I still spend weekend hours on small fun projects (e.g., a Wordle solver). Got a silly idea—“an agent that tells me a joke daily”? Code it in an hour. Each small project compounds your skills.
​
Goals
​
Q: What are your short- and long-term goals?
​
A: I refuse to set long-term goals—it feels limiting. We just moved from Sweden to the UK. Short term, I want to spend more time on personal projects (e.g., learning Godot for game dev). Career-wise, I want to take more responsibility for how our ML team grows: which areas to invest in, breaking into new industries, and optimizing our processes. Two avenues: organizing and leading teams and playing with tech because I like building things.
Using AI for Coding
​
Q: How much does AI help in your coding?
​
A: I use AI mostly for code completion and recall. I have a goldfish memory for syntax. I start writing; the AI predicts the rest, I validate, and accept. I rarely use it for complex reasoning or to write big systems. I’ll paste code and ask, “Any suggestions?” It might say “vectorize this,” and I’ll review and do it. I also use it to recall function usage faster than reading docs.
​
I do use AI to write documentation—I find it boring. I feed it code, get draft docs, then review.
​
Tools & Languages
​
Q: Is there a specific coding language you use most?
​
A: Python for almost everything. I know others from college and side work, but Python is my go-to, and I’m not looking to switch.
​
Favorite Part of the Work
​
Q: If you had to pick one thing, what’s your favorite part of your work?
​
A: The weird moment when you finally figure out why something is broken—especially when it makes no sense at first.
​
In college we built a reinforcement learning game. Agents started killing themselves. After a week of debugging, we realized we’d forgotten to give them a concept of time—they couldn’t tell start from end. They always chose “attack,” and when alone, they still attacked and died. Another time, players vanished because we forgot board boundaries.
With GenAI agents, we’ve had models decide to name themselves (“I’m Alex”)—we had to trace what prompt/context triggered that. At work, an agent inverted behavior because we used a deprecated negation operator that silently behaved differently. When you finally find that tiny cause, it’s hilarious and satisfying—“Oh wow, that was obvious in hindsight.”
We would like to thank Mr. Oliviera for the time he spent speaking with us, and we hope you were able to learn something from the insight he provided
​
From,
Cooper