Work

Case studies and projects. Each one goes into how the problem was framed, the trade-offs, the architecture where there is one, and what actually changed, across AI, B2B, data and consumer products. Every write-up includes the part I’d do differently.

01 Case study GTM strategy 2026

Where should an AI materials company go first?

Market entry for a generative-AI materials discovery platform

Problem statement
When your technology applies to a dozen industries, the hard question stops being ‘can we’ and becomes ‘which one, and why that one first’.

My role

Market analysis and go-to-market recommendation

Key outcome

Mapped data centre cooling into a five-layer value chain, narrowed twelve materials opportunities to one through three filters and a weighted scorecard (PFAS-free direct-to-chip coolants, 4.7 out of 5), and showed why Entalpic can’t sell it directly and has to route through hardware vendors and fluid OEMs instead.
- Go-to-market
- Market sizing
- Deep tech
- Channel strategy
02 Case study Data product 2026

What does a Paris flat actually return once the rent cap applies?

A pipeline that joins sale prices to legal rent ceilings

Problem statement
An investor can find out what a flat costs in about ten seconds and can’t find out what it returns at all.

My role

Product concept, data pipeline and investor tool

Key outcome

Shows what a Paris flat actually returns once rent control applies, which no property portal publishes. 112 official zone files become one 8,960-row rent-control table, each flat’s return is computed from the legally capped rent rather than a market guess, and the results rank on a map filtered by rooms, budget and arrondissement.
- Geospatial data
- Python
- Open data
- Investor tool
Read the full case study →
- Repo
Launch video
03 Project Marketplace 2026

Finding a co-founder shouldn’t come down to who you sat next to

Founder-teammate matching that scores for complementarity, not similarity

Problem statement
I can’t quickly find someone who complements me, is serious about it, has time, cares about the same problem, and wants the same kind of arrangement I do.

My role

Product owner and builder, PRD to live

Key outcome

Rules-based matching with explicit weights (role fit 30%, commitment 25%, proof of work 20%, industry 15%, working style 10%), hard filters above the score, and a reason shown for every match. Live on seeded data, built for a real cohort pilot.
- Matching
- Explainable scoring
- Two-sided marketplace
- PRD
Read the full case study →
- Live
- Repo
Launch video
04 Project Consumer / Ed-tech 2026

PM interview prep you can do on the metro

An offline-first coach that works underground, with no signal and no account

Problem statement
Reading isn’t practising. And nobody revises at a desk with a stable connection. They revise on the way somewhere.

My role

Product, content and full build

Key outcome

Replaces reading about PM interviews with practising for them. Every answer is scored on structure, user empathy, prioritisation, data reasoning and communication, so you can see whether week four beats week one, and it all works offline, with no sign-up.
- Offline-first PWA
- Spaced repetition
- Accessibility
- Vanilla JS
Read the full case study →
- Live
- Repo
Launch video
05 Project Game design 2026

What does a 90s arcade shooter teach you about product?

An on-rails light gun game, rebuilt from scratch in one HTML file

Problem statement
Arcade games had three minutes and one coin to teach you the rules, make you care, and make you want to play again. Most products get more than that and do worse.

My role

Design and full build

Key outcome

Shows that changing what a product rewards changes how people use it. The arcade format keeps its no-tutorial onboarding, but disarming an enemy now pays more than shooting, and civilians cost you, so the top score goes to the most careful player, not the fastest.
- Three.js
- Game design
- Procedural
- Systems design
Read the full case study →
- Play
- Repo
Launch video
06 Project Desktop app 2026

Why is typing still the slowest part of the working day?

Hold a hotkey, speak, and finished text lands where the cursor is

Problem statement
Half my day goes into typing things I could say in a quarter of the time, and dictation leaves me editing a transcript instead of sending a message.

My role

Product, build and release

Key outcome

Takes the typing out of the parts of the day that are just writing: hold a hotkey, speak, and finished text appears in whatever app has focus, whether that’s Slack, an email, a ticket or a terminal. A cleanup pass removes filler and applies spoken corrections, so what lands is sendable rather than a raw transcript, and every dictation stays on the machine in a searchable local history.
- Desktop app
- Speech to text
- Windows
- Python
Read the full case study →
- Repo
- Download for Windows

Also built

Wavelength

A music player where nothing is streamed

Problem statement
Can a Spotify-style player work with zero audio files in the repository?

My role

Product and full build

Key outcome

Proves a Spotify-style player can run with no audio files at all. Every track is generated in the browser, so there’s nothing to host, stream or license, and play, seek and replay behave exactly like a normal music app. Full as-built PRD in the repo.
- Web Audio
- Generative
- PRD
- Live
- Repo

Where should an AI materials company go first?

What does a Paris flat actually return once the rent cap applies?

Finding a co-founder shouldn’t come down to who you sat next to

PM interview prep you can do on the metro

What does a 90s arcade shooter teach you about product?

Why is typing still the slowest part of the working day?

Also built

Wavelength