NEW GAME SOLUTIONS  ·  PRACTICAL SYSTEMS  ·  NO JARGON  ·  NO BLOAT  ·  JUST TOOLS THAT WORK
NEW GAME SOLUTIONS → THE WORKBENCH
THE WORKBENCH

THE
WORKBENCH

A proving ground. Things I built to see if I could.

HOW TO READ THIS PAGE
Self-initiated builds, made to learn.
Data is synthetic, public or my own.
Status is labelled honestly: Built, Prototype or In progress.

THE BUILDS

PORTFOLIO
BUILT

Daily ROM Games

The problem
How far does AI-assisted development go when it runs on specification and verification instead of blind trust?
What it does
Four small games, shipped at dailyrom.games.
A build brief for each game before any code existed: several thousand words, every constant and rule.
Automated checks per game: puzzles provably solvable before they ship, and a build that fails if any colour pairing misses accessibility contrast standards.
Built with
AI-assisted development, written build briefs, automated build checks, row-level security, SQL
What it shows
Governing AI-built software when nobody is checking my homework.
Data
Own project. Play data from dailyrom.games.
Date built
September 2026
Not a pivot. Not a startup. It makes no money and the player numbers are small. That was never the point.
VISIT DAILYROM.GAMES →
PORTFOLIO
BUILT

Rollcall

The problem
Looking up a school should take ten seconds. On most school directories it takes a lot longer. The address is somewhere under the banner ads, the pop-up and the “sponsored results”. The official sources are accurate but spread across several government sites, each with its own search.
What it does
Search any Victorian school by name, suburb or postcode. Each of the 2,300 or so open schools gets one plain page. It shows the address, phone number, sector, council area and a local map. It also lists the nearest primary, secondary and specialist schools.
Suburb pages add Census 2021 context: population, median age, household income, children of school age, and the SEIFA disadvantage index with a decile legend. Council area pages list every school in the area.
Rollcall holds no kindergarten, zoning, NAPLAN or enrolment data. It links straight out to Find My School, My School, Find a Kinder, the VRQA State Register and the Schools Directory instead. The figures you read are the official ones.
Built with
Plain JavaScript and one shared template module. A Node script builds about 3,350 static pages. Hosted on Netlify. No backend, no accounts and no tracking.
What it shows
Pulling five public datasets into one clean model, and working out which licences allow what.
SEO at scale, with an automated audit for titles, canonical links, sitemaps and structured data.
Restraint: a page that answers the question and then stops. If advertising arrives, it will be one labelled ad at the foot of the page, never inside the facts.
Accessibility groundwork (semantic markup, skip link, keyboard-operable search), not yet formally audited.
Data
Real and public. DataVic School Locations 2026 (CC BY 4.0), ABS Census 2021 and SEIFA 2021 (CC BY 4.0), and ACARA identifiers for My School links.
Rollcall is an independent site and is not affiliated with any government body.
Status
Live at rollcallvic.com.au.
VISIT ROLLCALLVIC.COM.AU →
PORTFOLIO
BUILT

Secret Static

The problem
Can a screen show a message that your eyes miss but a phone camera reads instantly?
What it does
A browser toy that hides messages on a screen so a phone camera can read them while the people in the room cannot.
Four modes: Play (blue flicker and a party word game), Lab (twelve test cells), Slow Glow and Light Painter.
It is a toy, not security. Anyone with a camera can read the message.
Built with
One HTML file using WebGL. Built with Claude as an AI pair, then put through automated QA.
What it shows
Following the data, not the obvious idea. Visual patterns failed in the first round of testing, so timing became the method. Build, test and iterate on a real monitor and a real phone, with results logged each round.
Data
No uploads. Player names and scores are saved only on the device.
Test the assumption early, and let the evidence pick the direction.
TRY SECRET STATIC →
PORTFOLIO
PROTOTYPE

Daybook

The problem
Time gets written up late, from memory. Time that never gets logged never gets billed.
What it does
Paste a day or week of calendar events. Daybook turns them into draft time entries.
Client, matter, activity and description are assigned per entry. Durations round to a set billing increment, and raw time sits next to billed time.
Confirmed entries export to CSV for Xero or similar. Exported entries lock, so the same time cannot be billed twice.
Built with
Plain JavaScript, one HTML file, mock data in the browser. Planned backend: Supabase (Postgres, Sydney region).
What it shows
Scoping a narrow tool in a crowded market. Data modelling. Accessibility testing: WCAG 2.2 AA pass, axe-core clean, not yet tested with a screen reader.
Data
Synthetic. Sample clients and a sample calendar only.
Date built
October 2026
TRY THE PROTOTYPE →
PORTFOLIO
PROTOTYPE

The AI Spend and Value Tracker

The problem
Companies buy AI seats quickly and review them slowly, and a usage report shows activity, not value.
What it does
For each AI tool in a made-up 80-person company, it gives one answer: keep, right-size, consolidate, cut or investigate.
It says how sure it is, shows the working, and states what the method cannot tell you.
Value is a range of labelled guesses, never one return number. Sliders change the guesses and the answers move.
Built with
Plain HTML, CSS and JavaScript with hand-built SVG charts. Built with Claude Code to goal-first briefs I wrote and approved at each gate.
What it shows
Turning a messy question into a decision with a stated level of confidence. Testing AI-built analysis and reporting what failed, including my own errors and the AI's.
Data
Synthetic. Fictional company, illustrative prices.
Date built
October 2026
The point is the method, the testing and what failed.
TRY THE PROTOTYPE →
IN PROGRESS. STILL ON THE BENCH.
PORTFOLIO
IN PROGRESS

The AI Skill Library

The problem
Repetitive business tasks get handed to AI on faith, with no evidence that it works.
What it does
5 or 6 repetitive tasks a product business has: monthly reconciliation, product description drafts, support ticket triage, weekly sales summaries.
Each packaged as a reusable Claude Skill with a test set that scores the output. Evals rerun when a new model version ships, and what drifted gets documented.
Each gets a verdict: AI is good here, AI with a human check, or use a spreadsheet or a phone call instead.
Built with
Claude Skills, scored test sets
What it shows
Pass rates, failures and a recommendation, including when not to automate.
Data
Synthetic
Date built
In progress
IN PROGRESS. STILL ON THE BENCH.
PORTFOLIO
IN PROGRESS

The Enablement Field Kit

The problem
Polite rooms hide problems, and training that changes nothing afterwards is wasted time.
What it does
A workshop kit: facilitator guide, hands-on exercise, a short guidelines one-pager written to be read, and a 1:1 question log that feeds the next session's plan.
The plan is to publish a small, honest result, including anyone who preferred their old way.
Built with
Plain documents and a question log
What it shows
Training designed to change how people work afterwards, reported honestly.
Data
None collected yet.
Date built
In progress
Looking at my work as a prospective employer? Find me on LinkedIn.