Bridging Physics and Fintech: My Data Science Internship at ANNA by Carolina Bernandes

Over the past three months, I've had the opportunity to work as a Data Science Intern in the Business Admin team at ANNA, with Nick Turusin and Liev Garcia.

I was struck by the trust and ownership I was given straight away. Right from the start I was involved in end-to-end projects, from conducting exploratory analysis in BigQuery to shipping production deployments for our customers. Working across three very different projects, the common thread was always the intersection of data and product. My goal remained the same throughout: turning what we know about our customers into features that actually work for them.

My Background

With a background in physics, I’ve spent years learning to model complex systems and extract measurable outcomes from imperfect data. Throughout my academic trajectory, I’ve been drawn to the computational side of science – in particular machine learning and data-driven problem-solving.

Through research experiences in quantum optics, positronium spectroscopy, and data analysis for the LHCb experiment at CERN, I got used to working with large, noisy datasets. Aside from helping me develop technical skills, this also made me curious about applying analytical tools beyond academia, in a way that directly shaped real products and was applicable to everyday life.

Ultimately, joining as a Data Science Intern gave me the chance to translate a physics-trained mindset into industry-facing work, moving away from theoretical models and experiments and towards customer behaviour, product decisions, and production systems. Over the past three months, this transition has been both challenging and incredibly rewarding.

ANNA Year Wrapped: Data Storytelling in Action

The first project I worked on was "ANNA Year Wrapped", which consisted in presenting a summary of our customer’s year with the company: their spending patterns, earnings, and growth over the past twelve months.

My role was building the data collection and logic behind these profiles. This meant querying large-scale transaction databases, aggregating patterns across different categories, and incorporating that into a meaningful narrative.

My main focus was on striking a balance between finding the interesting trends hidden within common business transactions while ensuring the data remained an authentic, unmasked (albeit completely anonymised) reflection of the customer profile.

+Taxes Activation Agent: Orchestrating Production AI

The second project moved me closer to production AI systems. I built an LLM-powered agent to identify and engage customers who hadn't yet subscribed to ANNA's +Taxes plan.

This agent analyses each company's specific context – their business type, trading status, incorporation date – and presents personalised benefits and reasons for upgrading. I integrated this into a "Smart Checklist" within the mobile app, which ranked each item dynamically using an LLM to decide the relevance of each item to that particular user.

Building this meant working across the stack: from Python handlers to event-driven mobile app flows, alongside coordination with the mobile team to make sure there was a good customer experience.

Regional LLM for ANNA Australia – Transaction Categorisation

The third project I was involved in was the implementation of LLM transaction categorisation for ANNA Australia.

Because Australian financial data follows different rules and lacks the third-party context we rely on in the UK, I had to recalibrate our methodology to fit these new requirements. This included adapting our prompt engineering to handle Australian tax categories, designing new logic to work without Companies House data, and switching to Google Gemini 2.0 to optimise for regional performance and cost-efficiency.

Key Takeaways: Beyond the Model

Each project pushed me in a different direction, but they reinforced that the impact of data science relies mostly on connecting analysis to action.

Technically, I gained much more experience than I could have anticipated: LLM orchestration; data engineering from BigQuery through to production; and the infrastructure layer (Ansible, Docker, migrations) that underpins everything else.

Collaborating with different teams also reminded me that none of this work happens in isolation. It showed me how much of data science is communication – explaining technical constraints clearly, understanding product requirements from different perspectives, and building solutions that work for everyone involved.

The last three months of daily meetings, debugging sessions, and dealing with real customer data gave me a great sense of what it means to be a data scientist in a fast-paced environment. This experience was made all the more rewarding by the team’s constant willingness to help, which turned every technical challenge into a learning opportunity.

Getting the most out of your internship at ANNA

So, what can you do to make the most of an internship at ANNA? It helps to lean into the responsibility you’re given and use it as a valuable learning opportunity by asking questions early, sharing ideas openly, and getting comfortable with the product-led environment. Some of the most significant growth comes from talking to people across teams and understanding not just how something is built, but why it matters for customers.

Being proactive, curious and willing to participate can turn the internship from a learning exercise into a genuinely impactful experience.