Projects

What I learned

Deepened my understanding of deep learning maths by deriving and implementing backpropagation, Adam optimizer equations and the reparameterisation trick from first principles. Learned how to train generative models effectively—using free-bits KL divergence penalties and per-batch beta annealing to prevent posterior collapse in VAEs. Also gained substantial experience designing clean, idiomatic Rust library APIs for crates.io publication.

Challenges

Balancing reconstruction loss with KL divergence regularisation in VAEs without standard ML frameworks. Preventing posterior collapse required careful implementation of free-bits KL and per-batch annealing. In Rust, managing ownership, borrowing, and cache-friendly memory layouts for tensor operations without compromising execution speed was a rewarding systems challenge.

What I learned

This is my first venture into Rust and wanted to tackle the challenge head on. Learned many new ways of algorithmically solving a Sudoku.

Challenges

Getting file handling to work in my first Rust project was a struggle, as well as restructing the entire repo for better future use.

What I learned

Heuristics got me 80% of the way surprisingly fast. The LLM upgrade was less about accuracy and more about message quality - the basic version was correct but robotic.

Challenges

Getting the LLM to produce concise, conventional-commit-style messages without hallucinating context that wasn't in the diff. Originally I had a example in the LLM prompt. The LLM, however, just returned this example every time.

What I learned

Everything in this site was a new experience to me. I learned about API keys through the Spotify widget. This was also my first introduction to Tailwind and Next.js, on top of this hosting the site through vercel and connecting to my remote Github repo was a new challenge. I've learned a great deal of important and useful web skills through this project.

Challenges

Maintaining a consistent design system across components without a UI library. Choosing what elements go where whilst keeping a user's journey in mind. The challenge of making the site look good on both desktop and mobile was harder than I originally thought.

What I learned

Electron's split between main and renderer processes forces you to think carefully about where logic lives. Writing tests for a desktop app with real hardware dependencies taught me more about mocking strategy than any course. As the Product Manager, I was in the customer facing making and leading 5 sprint meetings with the team and Malin the duration of the project, this taught me time management skills and reinforced my presentation skills by clearly articulating our monthly work.

Challenges

Coordinating a team of five across a shared Electron/React codebase without stepping on each other - merge conflicts were a constant. Getting Electron packaging right, from icon resolution to app path handling, took longer than expected for something that felt like it should be trivial.

What I learned

Working with the RockYou dataset exposed how predictable human password choices really are. Feature engineering for text-based ML is harder than it looks - defining what makes a pattern took longer than building the model itself.

Challenges

Balancing recall vs precision for weak password detection. If I rebuilt this I'd spend more time on the feature pipeline before touching the model.

What I learned

Having never built a Chrome extension before, I learned how Manifest V3 works - including content scripts, message passing between the popup and page context, and securely handling API keys in a client-side environment. I also got hands-on experience with the Obsidian Local REST API and how to integrate it with a browser extension.

Challenges

The trickiest part was getting Chrome to communicate with Obsidian's local HTTPS server - the self-signed certificate caused fetch requests to fail silently until I traced the issue and configured Chrome to trust the local origin. Scraping the artifact content from Claude's React-rendered DOM also required careful inspection to find stable selectors, and converting the rendered HTML back to clean markdown with Turndown needed custom rules to preserve tables and heading styles.

What I learned

Django's ORM and auth system get you a long way fast - user authentication, profile management, and relational data modelling were all far less painful than building them from scratch would have been.

Challenges

Designing the data model to support both city rankings and a social follow graph cleanly. Getting the feed logic right - filtering posts by followed users while maintaining a coherent ranking system - required more careful query design than expected.