You hold a Master’s in Physics and Astrophysics. How does your background play into your work in data science and AI engineering?
Physics taught me two things that I lean on on a regular basis: easy methods to stay calm after I don’t know what’s happening, and easy methods to break a scary problem into smaller pieces until it’s not scary. Also… physics really humbles you. You learn fast that being “clever” doesn’t matter if you happen to can’t explain your considering or reproduce your results. That mindset might be probably the most useful thing I carried into data science and engineering.
You lately wrote a deep dive into your transition from an information scientist to an AI engineer. In your each day work at GLS, what’s the only biggest difference in mindset between these two roles?
For me, the largest shift was going from “Is that this model good?” to “Can this method survive real life?” Being an AI Engineer just isn’t a lot concerning the perfect answer but more about constructing something dependable. And truthfully, that change was uncomfortable at first… but it surely made my work feel far more useful.
You noted that while an information scientist might spend weeks tuning a model, an AI Engineer may need only three days to deploy it. How do you balance optimization with speed?
If we have now three days, I’m not chasing tiny improvements. I’m chasing confidence and reliability. So I’ll concentrate on a solid baseline that already works and on a straightforward technique to monitor what happens after launch.
I also like shipping in small steps. As a substitute of considering “deploy the ultimate thing,” I feel “deploy the smallest version that creates value without causing chaos.”
How do you think that we could use LLMs to bridge the gap between data scientists and DevOps? Are you able to share an example where this worked well for you?
Data scientists speak in experiments and results while DevOps folks speak in reliability and repeatability. I feel LLMs may help as a translator in a practical way. For example, to generate tests and documentation so what works on my machine becomes “it really works in production.”
A straightforward example from my very own work: after I’m constructing something like an API endpoint or a processing pipeline, I’ll use an LLM to assist draft the boring but necessary parts, like test cases, edge cases, and clear error messages. This accelerates the method rather a lot and keeps the motivation ongoing. I feel the bottom line is to treat the LLM as a junior who’s fast, helpful, and sometimes fallacious, so reviewing all the pieces is vital.
You’ve cited research suggesting an enormous growth in AI roles by 2027. If a junior data scientist could only learn one engineering skill this yr to remain competitive, what should or not it’s?
If I had to choose one, it might be to learn easy methods to ship your work in a repeatable way! Take one project and make it something that may run reliably without you babysitting it. Because in the actual world, one of the best model is useless if no one can use it. And the individuals who stand out are those who can take an idea from a notebook to something real.
Your recent work has focused heavily on LLMs and time series. Looking ahead into 2026, what’s the one emerging AI topic that you just are most excited to jot down about next?
I’m leaning increasingly toward writing about practical AI workflows (the way you go from an idea to something reliable). Besides, if I do write a couple of “hot” topic, I need it to be useful, not only exciting. I need to jot down about what works, what breaks… The world of knowledge science and AI is stuffed with tradeoffs and ambiguity, and that has been fascinating me rather a lot.
I’m also getting more inquisitive about AI as a system: how different pieces interact together… stay tuned for this years’ articles!
To learn more about Sara’s work and stay up-to-date together with her latest articles, you possibly can follow her on TDS or LinkedIn.
