Lab — software case studies

Things I build

By day I run machines at a precast concrete plant. Behind the lens I photograph architecture. And after hours I design, build and ship software — solo, end-to-end, with AI as a working partner the whole way. Below are seven case studies I'd put in front of a hiring manager: the problem, what I built, my role, the stack, and where it landed. The latest one went to the AMD Developer Hackathon.

Hiring? Start with the evidence page → every claim links to proof

20+ projects shipped 10 live products 6 open-source Solo · AI-native Web · Mobile · CLI · ML · Agent tooling

Selected case studies

Latest — AMD Developer Hackathon: ACT II

MixMind

Hackathon submission · Gemma 4 12B / LoRA / ROCm / Python

A plant-private concrete-mix AI — the R&D department a small plant can't afford to hire, running entirely on its own AMD GPU.

R² 0.910strength model, held-out

−17.3% CO₂on the approved mix

0 bytesrecipe data off-site

287 MBLoRA adapter, tuned on MI300X

Problem

Concrete is ~8% of global CO₂, and the mix-design know-how that could cut it sits inside a handful of corporate labs. Small plants can't hire an R&D department — and they will never send their recipes to a third-party API. So the AI has to live in the building, on hardware a plant can actually afford.

What I built

Two systems on one self-hosted, LoRA-fine-tuned Gemma 4 12B running on an AMD GPU: a Floor Copilot that answers plant-floor questions from the plant's own notes — with citations, and a refusal when the notes don't cover it — and a Mix Committee: four AI specialists that debate a recipe change grounded in a real strength model, then return minutes and a verdict. It rejects the cheapest, greenest mix when early strength fails the demould spec — judgment, not a yes-man.

My role

Solo, as a machine operator at a North American precast plant — fine-tune (LoRA on MI300X), retrieval design, the strength model, committee orchestration, UI and the submission.

Stack

Gemma 4 12B + LoRA served locally on AMD (ROCm); Python; gradient-boosting strength model on the UCI concrete dataset; retrieval-grounded generation.

AI workflow

Evaluated the fine-tune with a blind A/B LLM-judge from an independent model family, which preferred the tuned answers 2 : 1. Built AI-paired end-to-end.

Result

Working system with measured numbers (R² 0.910, −17.3% CO₂ on the approved mix); submitted to the AMD Developer Hackathon: ACT II on lablab.ai.

Repo ↗

Agentic Sentry

Open source · Bash / macOS

A runtime safety guard that watches what local AI coding agents actually do — and stops the dangerous stuff.

Problem

AI coding agents run shell commands on your machine with real power — rm -rf, reading SSH keys, sudo, network changes. One bad suggestion can wipe a directory or leak secrets, and there was no lightweight local guardrail to catch it.

What I built

A macOS runtime guard that monitors commands in real time. The default soft-block mode warns or blocks dangerous behaviour (deletes outside allowed directories, SSH-key access, sudo, network changes); an opt-in hard-enforcement mode adds emergency network/process kill-switches. Ships with audit logs and a recovery module.

My role

Solo — threat model, rule engine, shell hooks, install & recovery scripts, test harness, docs, and Homebrew packaging.

Stack

Bash, macOS shell hooks, structured safety-rule files, Homebrew tap, scripted test suite.

AI workflow

A guardrail for AI agents, built with AI: designed the rule model and enforcement logic AI-paired, with a test written for every safety rule.

Result

Open-source and installable via a Homebrew tap; documented with ARCHITECTURE and SECURITY guides.

Repo ↗ Homebrew tap ↗

AgentPay

Open source · TypeScript

A spend-policy gateway that lets AI agents pay for things — without ever holding your wallet.

Problem

As agents start paying for APIs and services across different rails (x402-style paid APIs, Alipay/WeChat-style payments), every company hits the same question: how do you let an agent spend money without handing it your credentials and an unlimited budget?

What I built

A cross-rail gateway that sits between agents and anything that charges them. Agents call paid resources through it; it enforces multi-currency budgets, per-transaction limits, payee allowlists and rail routing, and writes a full audit log. Agents never hold payment credentials.

My role

Solo — architecture, the spend-policy model, the rail abstraction, gateway implementation and tests.

Stack

TypeScript (zero-runtime-dependency core), Node, Vitest for the test suite.

AI workflow

Spec'd the policy model and rail abstraction with an AI pair, then drove it test-first so every policy rule is covered.

Result

Open-source working gateway with a green test suite; a reference for the emerging agent-payments problem.

Repo ↗

CloakPII

Open source · Python

Move customer data across borders — masked, encrypted and audit-ready — in one command.

Problem

When data leaves China or Singapore you're on the hook for PIPL / PDPA: personal data must be desensitized, the transfer encrypted, and a compliance report filed. Doing it by hand means stitching together PII detection, masking, AES encryption and a compliance checklist — and getting every one of them right.

What I built

A single Python CLI that takes a folder of files, masks (or reversibly tokenizes) the PII inside, encrypts every file with AES-256-GCM, and generates the PIPL / PDPA compliance report — so cross-border data is safe to move and there's an audit trail to prove it. Optional ML-based PII detection.

My role

Solo — CLI design, the masking/tokenization scheme, the crypto pipeline, the compliance-report mapping and packaging.

Stack

Python, the cryptography library (AES-256-GCM), PyYAML; optional ML PII detection; pip-installable.

AI workflow

AI-paired the detection rules and the mask-then-encrypt pipeline, and iterated the regulation-to-report mapping.

Result

Open-source CLI, pip-installable, with ROADMAP and project-summary docs.

Repo ↗

geo-engine

Open source · Claude Code skill

A repeatable engagement for getting a brand cited and recommended by AI answer engines.

Problem

When someone asks ChatGPT or Perplexity "best architectural photographer in Vancouver," a brand is either named or invisible. Most "GEO checklists" miss the real lever — off-site citations and a clean entity, not meta-tag tweaks.

What I built

An opinionated Claude Code skill that runs a full GEO/AEO engagement end-to-end: scope → audit → on-site (AI-crawler access, llms.txt, structured data, answer-first content) → off-site citation building → measurement and reporting — distilled from real builds into something you can run on your own sites or for clients.

My role

Solo — the methodology, the skill itself, the reference library and the client-deliverable templates.

Stack

A Claude Code skill (structured Markdown + references); schema.org, llms.txt and structured-data patterns.

AI workflow

It is an AI workflow — a packaged skill an AI agent executes end-to-end. Built and battle-tested on this very site.

Result

Open-source; the methodology behind this site's own GEO build.

Repo ↗

CanPay Insights

Live · React / TypeScript

A clean Canadian take-home-pay calculator, built to rank for the searches people actually type.

Problem

Canadians struggle to work out real take-home pay — federal and provincial tax, CPP, EI, overtime — and most calculators are cluttered, province-blind, or both.

What I built

A fast pay calculator covering every province (federal & provincial tax, CPP, EI, hourly / salary / overtime, timesheet tools), built SEO-first and shipped on its own domain — with a companion Android app.

My role

Solo — product, UI, the calculation engine, SEO and deploy, plus the React Native mobile build.

Stack

React, TypeScript, Vite, Tailwind, Gemini (@google/genai); deployed on Vercel at a custom domain.

AI workflow

Scaffolded in Google AI Studio (Gemini), then hand-extended into a production calculator and an SEO-built site.

Result

Live at canpayinsights.ca, with an Android companion app (CanPay App).

Live demo ↗ iOS App ↗ Repo ↗

Delta Plant

Live · React / TypeScript + Supabase

The tool the plant floor actually uses to manage brick & paver mix recipes.

Problem

A brick and paver plant's production recipes — mix formulas, colours, sizes — lived in spreadsheets and people's heads. That's error-prone and hard to hand off between operators on different shifts.

What I built

A production recipe-management app for plant operators: create and manage mix formulas, colours and sizes, backed by a real database so the floor works from one source of truth.

My role

Solo — built for my own workplace: data model, UI, the Supabase backend and deploy, iterated with the operators using it.

Stack

React, TypeScript, Vite, React Router, Supabase; deployed on Vercel.

AI workflow

Prototyped in Google AI Studio, wired up to Supabase, then refined against real operator feedback.

Result

In production use by operators on the plant floor.

Live demo ↗ Private repo

More builds

FlashPick ↗ live

A macOS app for photographers — cull and pick your best frames fast. Distributed as a signed DMG.

Keepsake Guide ↗ live

Canada-wide funeral, cremation & obituary directory — Next.js, server-rendered, built for SEO.

SteadyQuant ↗ live

A beginner-friendly quantitative-investing tool — plain-language portfolio guidance.

IP Detective ↗ live

A cross-border e-commerce environment checker — IP, timezone and network fingerprint.

Show Notes Generator ↗ live

Turns a podcast or video transcript into structured show notes and timestamps.

Atlas of Us ↗ live

An interactive personal travel atlas — map, pin and remember everywhere you've been.

Prime Nexus ↗ live

An AI API-token shop — unified access to LLM APIs, web chat and tooling.

Low Battery Studio ↗ live

An indie merch brand — apparel and home goods for people running on 1%.

TikTok Shop US Playbook ↗ repo

A bilingual English / Chinese skill generating compliant TikTok Shop US seller playbooks & SOPs.

ShiftBuddy private

A shift-work scheduling & roster helper — plan shifts, track hours and hand off between operators.

VisaScout ↗ iOS App

A global mobility analyzer — visa requirements for any passport to any destination, with visa-free / e-visa / on-arrival breakdowns. Native iOS app.

CanadaPR private

A Canadian immigration information platform — programs, pathways and requirements.

TieTie private

Snap-a-receipt expense capture & approval — AI OCR, then review, sign-off and reimburse.

TangFei private

Luggage-allowance resale quoting — a mini-program plus a web back-office.

AU Alumni private

An alumni platform — member directory, digital alumni card and school news, mobile-first.

Paver Tycoon prototype

A tongue-in-cheek browser-game prototype — an experiment in game loops.

All of these were built solo, AI-paired. Private items are client or internal builds — happy to walk through any of them on request. Open-source repos and live demos are linked above.