Researcher, ML engineer, ponderer.

From wanting to study philosophy, I found mathematics - a compromise that ended up not really being a compromise. I found ML working at a nonprofit my freshman year: building quantitative tools for analysts working in anti-trafficking intelligence, and realising that the statistical methods I kept reaching for had a lot of assumptions I couldn't fully account for. So I taught myself what I needed, pitched a research project, won funding, and kept pulling on that thread.

That work left a mark on how I think about models. The data I was working with was shaped by language, dialect, and cultural context. A model trained without accounting for how different communities use language, or how social norms inflect the signals you're measuring, was just wrong in ways that mattered. Bias in these systems was an inherent design problem. Since this work preceded my formal education in the subject, it kinda defined the way I think about representation learning.

My BSc felt like putting vocabulary to things I'd been doing by feel. An internship that turned into a research role on a team working on AI assurance and formal methods got me interested in neuro-symbolic AI — and gave me an excuse to focus my dissertation on neuro-symbolic mathematical reasoning, which is really a question about the interface between formal correctness and learned approximation. At Two Six Technologies I work on hybrid symbolic-neural systems for high-assurance environments: knowledge extraction pipelines, automated reasoning, logic programming, entity resolution. Alongside that I've been doing independent interpretability work, and running a small business applying research to interesting problems and bridging creativity with technical capability. The question underneath it all is the same one: what are these systems actually doing, how do they effect people, and can we actually say something precise about it?

Something that doesn't fit neatly on a CV: I think there's a craftsmanship to technical work that gets squashed when the ethos/priority/approach is a "manufacturing" one — applying proven approaches toward what ends up being variations of familiar solutions. The choices about what to make explicit and what to leave implicit, the framing of what the problem actually is, the story a system tells about itself - these matter. I want to somehow change how people relate to technical subjects, math included. Rigour and aesthetic sensibility aren't contradictory, often the most responsible solution is also the most elegant one, and I think we lose something important when we stop believing that.

---

Experience

---

Intellectual interests

Mechanistic interpretability — what SAEs reveal about feature geometry in language models, and what that might tell us about compositionality and abstraction. AI bias and the question of what it would mean to take distributional harm seriously as a formal constraint rather than a post-hoc audit. The neuro-symbolic interface: not as a compromise between two paradigms but as a genuine theoretical question about the relationship between learned and explicit representations.

The philosophy of language sits underneath a lot of this — particularly questions about meaning, reference, and what it would take for a formal system to "understand" something rather than model it. Linguistics in the structural and cognitive tradition. Computational social science as a way of keeping the stakes in view. I also organise a speaker series bringing together researchers across interpretable AI, formal methods, network analysis, and neuroscience — partly because I've noticed these areas tend to meet eachother from different angles.

---

Now →

---