Gavin Wu  ·  Providence  ·  Auckland

I build things that fix something plainly broken.1

A grading tool for teachers who were drowning in marking. A volatility surface, because the clean version did not exist. A bottle cap that tells you the milk has gone off.

Gavin Wu

I study applied mathematics and computer science at Brown. I grew up in Auckland, and most of what is below started there — usually because something in front of me was obviously broken and nobody else wanted to own it.

GitHubLinkedIn

Things I have built

VolFoundry

Quantitative research

An arbitrage-checked implied volatility surface for crypto options. It pulls the live Deribit option chain, inverts each quote for implied volatility, calibrates SVI slice by slice, runs static-arbitrage checks and fits an SSVI surface across the whole grid.

The fitting is the easy half. The interesting half is deciding what to do when the market quotes something that cannot be arbitrage-free. It does, most days, and a surface that just shrugs is worse than no surface.

Live surfaceSource

maturitylog-moneyness
SSVI surface fitted across log-moneyness and maturity. Steep front-month smile flattening out along the term structure.

NCEAGrader

Founder · AI assessment

An AI grading and exam-authenticity platform used in New Zealand secondary schools. Sole developer. The core is a four-agent pipeline with strict parameter outputs, which brought backend API costs down 37.3% against the single-prompt version it replaced.

I have given eight talks on exam authenticity and NZQA reform across nine schools, and pitched the roadmap to Prime Minister Christopher Luxon and Education Minister Erica Stanford as a candidate framework for NZQA’s digital assessment tools.

nceagrader.com

Ingestscript, rubricParseparsed fieldsGradestrict paramsVerifyauthenticityBackend API cost−37.3%vs. single-prompt baseline
Four agents with strict parameter outputs. Splitting the work made each call smaller and cheaper than the monolithic prompt it replaced.

Enspire

Cofounder · Full-stack

A multi-region learning platform on Next.js, TypeScript, Python and MongoDB, serving more than 10,000 monthly active users across eight countries. I architected it, then spent most of my time on the unglamorous parts.

An AI summarisation pipeline with chunked retrieval and edge caching cut p95 cold-start latency by roughly a third. Technical SEO and multi-locale work moved the Lighthouse score from 65 to 98.

enspire.lol

8 regionsedge cacheoriginp95 cold startbaselineshipped−33%
Eight country regions hitting a shared edge cache. Chunked retrieval keeps the summariser off the critical path.

FluxCap

Founder · Materials

A bottle cap that tells you whether the milk has gone off. Fresh milk sits near pH 6.7; as lactic acid builds it falls toward 4.6, and a halochromic ring reads that shift from green through yellow to red.

No electronics, no app, nothing for anyone to remember to do. The CAD model is a two-shot injection-moulded HDPE housing with a sensor cartridge behind an ultrasonically welded ePTFE membrane. Costed to a NZ$74,600 pilot.

Chemistry paper

indicatorpH6.76.05.55.04.6freshspoiled
Lactic acid accumulates as milk spoils. The ring reads the pH drop directly, so there is nothing to scan.

JudgeGuard

Open source · Evaluation

Unit tests, regression tests and CI release gates for LLM-as-a-judge evaluators. Test the evaluator before you trust the evaluation.

Teams gate a release on an LLM judge and never test the judge itself. It can flip its verdict when you swap which answer is shown first, or disagree with itself on byte-identical input. JudgeGuard runs five deterministic probes against the evaluator and blocks the release on the first one that breaks.

Source

non-determinismposition biasmetadata leakageformat sensitivitymissed defectsrelease blocked on 1 of 5 probes
Five deterministic probes run as a gate, not a dashboard. This one ships with a judge that picks whichever answer goes first. The position-bias probe fails, and the release blocks.

Roughly in order

Current

Frontier Space Lab

Member of Technical Staff

Designing algorithms and agent pipelines to automate parts of the R&D workflow for lunar resource extraction.

2026

Icehouse Ventures

Technical Intern

Researched 2,400+ investment prospects and merged the outreach lists from five regions into one database. Built an automated LinkedIn research agent and deterministic/fuzzy-search gateways for the rest of the team.

Mar 2026 →

SkillsVR

Research and strategy, reporting to the CEO

Centralised and now manage a $100M cap table across 153 investors. Worked through three weeks of stalled payroll alongside the chief executive, including a one-on-one with a Malaysian senator.

Jan 2026 →

WakaForward

Strategy consultant and AI/ML analyst

Shipped a Python and AWS Textract pipeline that reads tax forms and bills of lading, working with five engineers in Chennai. Modelled regional demand with the CEO to design the APAC go-to-market.

Jun 2024 →

National Life Literacy Programme

Founder and director

New Zealand’s first free life-literacy curriculum, covering financial skills, CV writing and loan navigation, built as modules that drop into existing classes in low-decile schools. Peer-to-peer by design, so the first cohort of thirty became the ambassadors who carried it onward.

Jul 2021 →

Office of MP Dan Bidois

Political strategy and data analysis

Joined the office as the party’s youngest member. Helped launch the North Shore’s first homeless shelter by pulling together thirty-plus regional ministry leaders who had not been in the same room before.

Recognition

1st in Oceania

Global Blue Ocean Strategy Competition, run by Harvard Business Review and Morgan Stanley. Top 30 of 12,700+ competitors across 163 countries.

4th in the world

International Tertiary Ethics Olympiad. Captained New Zealand as the first under-18 ever to represent the country.

Nine NZQA Scholarships

Each awarded to the top 3% nationally in its subject, between 2023 and 2025.

NVIDIA Inception

Partner with $100,000 in API funding. Y Combinator Startup School, selected from more than 30,000 applicants.

Brown University

Applied mathematics and computer science, class of 2030, on full scholarship.

Three papers

On curriculum reform, spoilage chemistry and the structure of Pascal’s Wager. All with DOIs.

Google DevSprint Hackathon

1st place and MVP award among 111 participants from Asia, the Pacific and Europe. Built a logistic-regression classifier for client-sensitive data as team captain.