Build
These are arranged by what they prove: an independent founder build, systems that compound, and research that changed a product. The stack matters when it changes the product; the case studies show how. Every entry says what I personally owned and what I did not, because that is the part a reader cannot check for themselves.
IA founder build
-
Vaulth AI2026—
Who should hold a family's medical records — the hospital, the app, or the family? I architected the custody and consent platform that answers "the family," and shipped it as the sole engineer.
207 endpoints67 tables7 services 630 testspre-launch
Mine: all of it — product, UX, backend, web, mobile, infrastructure, AI, security and compliance. Not mine: nothing, which is the risk as much as the claim.
Health records custody · Founder & CTO · vaulthai.health · node · next.js · flutter · postgres · vertex ai
IISystems that compound
-
A prior platform portfolio2024—26
A previous product role across a multi-product games, payments, and document-AI portfolio. I owned the research that set the roadmap, the onboarding strategy, a shared-platform decision, and two pre-agreed feature kills.
20% faster first action12% D7 retention 15% feature adoption
portfolio strategy · shared platform design · research · delivery
-
Arthside project
Can a finance app understand your money without ever being told what it is? Bank SMS, PDF statements and CSV exports parsed on-device by a deterministic-first cascade, with a 4-bit Gemma-3 behind a hand-written FFI bridge for the long tail. Zero bytes transmitted.
0 bytes transmitted4-function native surface 11 decisionsphase 4 of 5 — awaiting real-use measurement
Mine: the whole thing, solo, unreleased. Not mine: the models — Gemma-3 quantised, llama.cpp, and the parsers below the AI tier.
On-device financial intelligence · llama.cpp · dart ffi · gbnf · flutter · kotlin · encrypted sqlite
-
Cricket Managerside project
How do you make a simulated cricketer behave like the real one — and how would you know if you had failed? A ball-by-ball engine calibrated on 2.07 million real T20 deliveries, player ratings adapted from the Earth Similarity Index, and a live event-sourced auction.
2.07M deliveries modelled10 decisions 0.98 correlation — rejected as headline external play remains untested
Mine: the simulation engine, the rating model and the decision to distrust it, the auction. Not mine: the ball-by-ball dataset.
simulation engine · statistical modelling · real-time multiplayer · event sourcing
IIIResearch that changed the product
-
Portfolio research2024—26
An interface inventory and diary studies that made a multi-product roadmap defensible. The work identified adoption blockers before a product-management role existed; the resulting onboarding redesign cut time-to-first-action by 20%.
interface inventory, whole portfoliodiary studies companion portfolio case
User base growth · research · onboarding · retention
-
iAuto Analytics2022—24
Qualified traffic arrived and left. Why? Funnel analysis and interviews found the answer in the landing pages, not the product.
A/B validateddrop-off located before explained
data-driven ux · a/b testing · hotjar · google marketing platform
-
PluxBox2022
How many steps can you remove from a no-code platform before it stops being usable? Thirty user interviews decided it, not opinion.
30+ interviews30% fewer setup steps 25% fewer support tickets
no-code saas · user research · onboarding
Foundations — university, 2020–22 Three early projects · where the habit of checking started
These are student projects and are filed as such. They are here because the method in them is the method still in use, not because the artefacts are current.
-
Survive the Crisis2020—22
Can a game measure how prepared someone is for a crisis better than a questionnaire can? Built for VRT Twente Safety Region to find out.
SUS 75.63full UEQ published4 limits, all published
serious game · vanilla javascript · ux research
-
Cashpool2020—22
Can four people from four disciplines validate an app before anyone builds it? A first-year thesis at UPM that ran the human-centred process end to end.
17 interviewscard sort settled the argument shipped against our own data
android · figma · card sorting · java · team of four
-
ANT2020—22
What do students actually feel during a lecture they can't keep up with? The design answered the emotion first and the note-taking second.
5 emotional failure modes mappedUEQ tested
automatic note taking · emotional design · figma
Ten projects, one recurring habit: find the real constraint before choosing the build.
Building something that should exist? Or want to argue with something here? Open to conversations.