2025 MSc Project Internship

Popsa sits at the intersection of generative AI, photography and eCommerce. According to Deloitte, we are one of the UK’s fastest-growing tech startups, and in 2022 the Financial Times ranked us in the Top 5 fastest growing software companies in the whole of Europe.

We have the backing of some excellent investors. Our native iOS and Android apps are available in 12 languages – attracting more than 5 million users to date – and we ship to 50 countries around the world. Only 15% of Popsa’s revenue comes from the UK (France is our top market whilst the US is our fastest growing).

? Read more about our journey so far

? Check out our Soho office in London

People come to us to create photobooks and other products, but what we really do is help them process their lives and relationships through reflection on their photos.

Our ambition is to be a space to explore all of life’s moments and the emotions that come with them – not just the highlights. So holidays, but also lazy days and sad days too. We believe the more you reflect, the better you understand yourself and your connections with others.

We’re developing an AI-driven platform that effortlessly generates personalised content based on the data contained in photos, removing countless anxieties usually associated with creative tasks and empowering everyday people with professional storytelling techniques for the first time.

? Read more about our vision for the future of Popsa

Founded in 2016, we’ve already built an award-winning app that’s made printing your memories so easy and accessible that anyone can do it. As a result, we receive 20,000 orders every week.

However, behind the scenes we’re building the next generation of automatic products – an order of magnitude more sophisticated than what you see today.

UCL “Machine Learning” MSc students

Popsa is providing the opportunity for a qualified UCL students to join our team for 8 weeks anytime between early June till September 2025.

We will be paying a stipend at London Living Wage (capped at 40 hours/week). 4-5 days will be working from our Soho office.

Proposed Summer Thesis Projects

Project 1 – Photo Library Knowledge Graph Prediction

Background

Dataset – A person’s photo library contains the places and people which are important to them

Link Property Prediction – Can we identify familiar people and locations from a graph of this data?

Labels – We have been asking users to label familiar places, people and relationships

Proposal

Build a knowledge graph – With nodes for photos, places, people & events; edges link photos with identified people, places & events

Hypothesis – A relationship can be inferred based on geo & temporal density of shared photos

Random Walk, Node2Vec – First algorithms to explore

Graph CNN / Transformer – May be required to achieve high accuracy

Project 2 – Aligning Edge-Based Vision Encoders with LLaMA

Background

Popsa has a privacy-first approach – No user photos are uploaded unless they are needed for printing

Large Multimodal Language Models – Large-scale models that embed images for advanced visual-text reasoning, but are expensive to run and memory hungry.

Edge-based Vision Encoders – Lightweight, on-device vision encoders that produce efficient embeddings with minimal resource usage.

Proposal

Data Generation – Extract embeddings from both an open-source Large Language Model and an edge-based vision encoder for millions of images.

Projection Layer Training – Learn a mapping from the vision encoder embeddings to the LLM embeddings

System Integration and Evaluation – Replace the LLMs heavy vision encoder with your projection during inference and evaluate the results.

Project 3 – Reinforcement Learning for Email Marketing

Background

Current email strategy – Emails are sent based on heuristic schedules as well as user activity

Available data – Historical email campaign data, user purchase history, user activity, email unsubscribe data

Objective – Create an email marketing framework that maximises user engagement and long-term value (LTV).

Proposal

Reinforcement learning (RL) – Train an RL agent to optimise email marketing using user activity, email send times and purchase activity as inputs.

Experiment with reward signals – Incorporating LTV, user engagement, email unsubscribes, etc.

Experiment with RL algorithms – Q-learning, Deep Q-Networks (DQN), Proximal Policy Optimisation (PPO), etc.

Develop framework for agent evaluation

Potential extensions – Recurrent RL, adaptive rewards, online learning, content optimisation, etc.

Requirements from your previous projects

• Project 1: Competence with Python & Understanding of how to encode information as a graph essential; Knowledge of graph machine learning algorithms & Experience working with databases desirable
• Project 2: Familiarity with machine learning fundamentals (loss functions, optimization) essential; Strong Python and PyTorch skills for neural network training & Experience handling moderate-to-large datasets and data pipelines desirable
• Project 3: Competence with Python & Understanding of core RL concepts essential; Awareness of advanced RL techniques & Experience working with databases desirable

Decision Timeline

• 7 Feb – Applications Close
• Before end of Feb – Decision

London Living Wage

PI261372720