Are you a final-year master’s student looking for a thesis project that blends real-world impact with cutting-edge tech?

At Framna, we are passionate about creating digital products that shape markets and make everyday life better - and we want you to be part of it.

Every year, we welcome a group of master’s students into our Stockholm studio to spend their final semester with us. You will sit alongside our developers, designers, testers, analysts, and product delivery leads, get a place to sit in our cozy office, and join our social and learning activities - from lunch-and-learns to our weekly Friday demos and AWs. It is a chance to do your thesis in a collaborative setting, surrounded by people who are happy to share what they know.

Background

Large frontier models are powerful, but they come with real trade-offs: cost, latency, privacy concerns, and energy consumption. Meanwhile, smaller open-weight models have become surprisingly capable when fine-tuned or prompted for specific tasks.

The question is not simply which model is better - but when smaller models are good enough, and whether a smart combination of small and large models can get the best of both worlds.

In this thesis project, you will design and evaluate a routing-based AI system that directs requests to the most appropriate model depending on task type and difficulty, and compare it against simply always using a frontier model or always using a smaller one.

What you will do

Some potential tasks within this thesis include:

• Defining a benchmark covering tasks such as summarisation, classification, question answering, text generation, and structured decision support

• Comparing three core approaches: a frontier model baseline, small specialised models, and a routing or cascading system that directs each request to the right model

• Evaluating the router against reference points including always-small, always-frontier, random routing, and an oracle router that selects the best model in hindsight

• Assessing each approach across dimensions such as output quality, accuracy, hallucination rate, latency, cost, energy use, and privacy constraints

• Exploring techniques such as prompt engineering, lightweight fine-tuning, retrieval-augmented generation, self-checking, or ensemble workflows

• Recommending when a small model is good enough, when a frontier model is needed, and whether routing between them is worth the added complexity

Who you are

This thesis may suit you if you:

• Are interested in AI systems design, model evaluation, and efficient inference

• Are curious about the practical trade-offs of deploying AI in real products

• Enjoy building evaluation frameworks and comparing model performance

• Plan to start your master's thesis project in January 2027

• Have the opportunity to work from our Stockholm studio

About Framna

Framna means ‘to bring forward’ in Old Norse. And that is exactly what we do. We partner with industry leaders (and those about to be) to create excellent digital products that define markets, reshape industries, and deliver meaningful impact.

Born from the union of digital agencies Bontouch, Move, and Shape, Framna emerged from a strong, product-led culture. Today, over 600 of us collaborate across ten studios in Denmark, the Netherlands, Poland, Sweden, Switzerland, and the US.

Every day, millions of users around the world rely on the products we help shape. From seamless payments to daily commutes, from better health management to smarter shopping. Our work supports more than €150 million in daily transactions and consistently earns an average App Store rating of 4.5.

Together with some of the world’s most ambitious brands, including Essity, SJ, Swish, and SEB, we craft products that lead markets and push boundaries. Because when products win, businesses win. We call it Win by product.

Our hiring process

• Psychometric assessment - personality and logic test

• Interview with our recruiter

• Case assignment

• Case interview