This summer the UnlikelyAI team welcomed 4 interns that multiplied the energy of the team with their fresh ideas, energy, drive, and full charge of the projects they delivered. Over three months, they became an essential part of the team. They developed features that went live, did fresh research into how LLMs behave, a Python package and a drawing-game rivalry the office still cannot settle.

Hands on projects accelerating the team

The four of them parachuted onto the projects that needed them most, accelerating the team with work that mattered. Here is what each intern took on and learnt along the way:

Charlot Eberlein built a redaction layer for one of the data pipelines for the UnlikelyAI Uncheck product, keeping personally identifiable information private without stripping away the context that makes it useful. Read more about her project in the blog she wrote.

"It feels great to have made a real, meaningful contribution. I've learnt a lot about how agile software development works in practice, which I'll take into my next year of university." - Charlot

Xietao Wang Lin went deep into research of iterative loops in LLMs, feeding a model's output back in as its next input. He also zeroed in on round-trip code translation and bug-fixing, because code is one of the few things you can actually mark right or wrong. His verdict on LLMs is written up in a paper he published.

"Turns out LLMs are incredibly messy, so it required quite a bit of tinkering. I also learned about many of the delicious snacks in the office, so much that I've started buying them for myself at home." - Xietao

Georgiy Savchenko built a Python package that automatically optimises LLM pipelines, with two optimisers doing the heavy lifting: one that builds a playbook of curated knowledge for approaching a problem, and one that writes the harness around the model itself. What used to take an engineer a week, now happens on its own.

"Designing a modular package around a technique that can literally change your codebase was an interesting challenge, and I had a lot of fun doing it." - Georgiy

Hassan Himaz went after two very important challenges to develop Uncheck. He built systems that break dense, compound regulatory requirements into smaller questions a reviewer can actually answer. He reworked the ingestion side so the system keeps track of the kind of text it is reading and exactly where each extracted fact came from.

"I've really enjoyed combining research with practical software engineering, testing ideas on real examples, learning from the failure cases and turning the strongest findings into usable features." - Hassan

And so much more they brought

Outside the impressive demos and hours spent writing code, they brought positivity, turned our office drawing game into something close to a professional sport and the kind of presence that makes an office better on a Monday morning.

They rounded things off with their project presentations, three months of real impact laid out in full. A cohort of absolute superstars, and living proof of what ownership can look like in early careers.

While we are sad to see them go, we know they have a bright future ahead of them. If this is what they can do in a single summer, the rest of the world is in for quite a lot.

Curious about work at UnlikelyAI? Register your interest on our careers page.