KEATH.ai
October 2023 – October 2025
While at university, I was also the first employee at KEATH. On the side of getting ChatGPT to quickly do my assignments, and on the side of socialising, I was working incredibly hard at this fun, new, exciting startup that we thought could go right to IPO.
It was genuinely super exciting for me, and I worked so hard that I missed a final assignment deadline by 2 weeks. I managed to convince the lecturer that she'd given me the wrong deadline rather than that I'd missed it — still got a capped 40% score in the end. I really don't suspect she's reading this, but apologies if you are.
Because the role was so generalist and self-directed, it often came down to me having intuition about what needed fixing, and then fixing it. That actually led me into user experience design as it was a massive bottleneck at the company. I learned I had pretty neat product intuition.
The core challenge was making software intuitive for lecturers of any age and technical competence. We achieved this, and I completely rebuilt the entire user interface on my own, straight into code that's being used today. Made the design system, revamped the whole frontend.
I personally believe the UX role will be impossible to separate from frontend engineering roles, as there's no point having both. I already often design straight into code and skip Figma just for brevity.
We've had thousands of teachers sign up and mark hundreds of thousands of assessments internationally, and partnered with 25 prestigious universities globally as well. That's something I'm super proud of.
I did hundreds of sales calls, and Saas demos with the product we made. This was incredible experience for me to learn how to speak the language of the prospect, let them come to their own conclusions, and how to sell something generally. Being the builder too meant I learnt things about and around the product I would'nt have. Often I'd develop something new straight off of a call. I like to think this iterative process was a big part in the success.
I then stepped in and helped with the fundraise — closed the $50M valuation round out of my Google Drive data room. Made the pitch decks, one pagers, and created and edited all marketing videos. We went into different countries and sold directly to institutions and governments, notably to El Salvador's Ministry of Education.
ELO — Spell Check for Contested Knowledge
July 2025 – Present
We started by benchmarking loads of different LLMs and their outputs, trying to understand what they're good and bad at. We built a smart router that would route your query to the best LLM for your request.
In an attempt to decide whether this was commercially viable, we modelled this to researchers — people who needed the best performing LLM so they didn't face hallucinations and got the best intelligence. And we found something interesting: different LLMs completely disagreed with each other on what seemed like basic science, especially to researchers who'd been studying that field their whole lives.
Here's what we realised: LLMs hallucinate when they run out of training data. That's causal, that's one-to-one. When you track where they're hallucinating against each other, that vector point where there isn't consensus is actually where the front of science is. It's where we don't know the answer — where there's contested knowledge in the field because it's no longer in the training data and the models are guessing.
We mapped this consensus voting mechanism across flagship LLMs and built something we were really excited about. We optimised it down into something small and scalable that post-processed written work and flagged either false or contested information in your document.
We called it a spell check for contested knowledge.
We believed this could be a massive system for truth generally as AI gets better and the cost of intelligence drops. Within science, so much of the work is knowing what we don't know — asking the right questions, chasing the correct targets. In pharmaceuticals, chasing the wrong targets is a huge problem. 90% of drugs fail, right?
So we implemented this into an all-in-one research paper writing application, AI-augmented to speed up the research cycle generally (administrative tasks take up like 50% of researcher time anyway). We built the MVP in two months, got researchers from Cambridge, Bristol, and Imperial testing it, and had early users publishing significantly faster.
Consulting — The Stuff In Between
2023 – Present
Throughout all of this, I've done consulting work. It's how I learned originally and how I've kept learning since.
Chatbots — Built curriculum-aware AI tutors with custom assessment workflows, deployed to actual students. The technical challenge was getting the RAG pipeline to use only the curriculum, and building a handoff system that introduced the teacher correctly. Also, building a more elongated tutoring experience where the answers were purposefully delayed and progressively revealed - matching real teaching methods. (ChatGPT and Gemini did this, but about 18 months later)
Fajj (Saudi Arabia) — Connected Saudi hotels with Umrah travel agents. Built outreach systems that qualified leads automatically, and booked 10+ warm calls a week. Enterprise B2B in a market I knew nothing about — useful exercise in figuring out new domains quickly.
The Turing Learning Initiative — Automated business operations, rebuilt their website, upskilled their dev team on AI tooling. Also developed course content for their AI proficiency certification, which forced me to articulate what I'd learned intuitively.
Various early-stage builds — AI chatbots, lead-gen systems, MVPs for trainer marketplaces, Nail Salons, Insolvency, solving their customer facing sites, and the systems behind it. AI chatbots for lead gen etc. The common pattern is companies that need to move fast and don't have the internal capability yet.
Consulting teaches you to solve problems quickly with incomplete context. You can't spend three months learning the codebase — you have forty hours and a deadline.
What I Learned
Frontend development is about constraints. The best interfaces aren't the ones with the most features — they're the ones where you've made the hard decisions about what to leave out.
Sales is just applied curiosity. The good calls were the ones where I asked questions instead of pitching. Universities have weird problems you'd never guess. It's easy to sell something good.
Systems beat hustle. Early at KEATH, I worked constantly but inefficiently. Later, I built automations and workflows that did the repeat work for me. I implement a rule with my work, and then with ELO, that we don't do an automatable task more than once.
Technical and non-technical is a false divide. I'm not an engineer. I can build products, read code, architect systems — but I'll never be the person to optimise database queries. The space between "can code" and "can't code" is where a lot of value gets created. Translating what's possible technically into what's valuable commercially. People usually only do one side of it.