Software & AI

Client delivery, our own products, and applied modelling — from keeping a production system healthy for years to turning industrial data into something a person can act on.

Client delivery

What a buyer should weigh is not whether a supplier can build something. It is whether they stay.

Ongoing since 2022

Content platform for a local in-person education provider

A public site and a purpose-built admin backend, on S3 and CloudFront, letting non-technical staff update schedules, faculty, policies and FAQs themselves.

The engineering story is the constraint set: content changes often, the people changing it are not developers, and the whole thing has to run cheaply and stay up. That points at a CMS backend behind static CDN delivery — which is, not coincidentally, the same reasoning behind the site you are reading.

Our own products

Open source · npm / PyPI / Homebrew

wtcraft

Git-native governance for AI coding agents. Multiple agents run in separate worktrees, each bound by a written task contract that states scope, off-limits areas and verification steps. State lives in git; there is no runtime and no database.

The point is that you can check it yourself — npx wtcraft init in any repository. In business software, evidence you can reproduce outweighs a description you have to take on faith.

wtcraft →

iOS · fully on-device

Wishes by Hand

A handwritten card app with Metal-backed low-latency pressure strokes, classical card designs and a calligraphy practice room. Everything runs on the device; nothing but iCloud sync leaves it.

The stroke scoring is geometric and rule-based, not a model. What it demonstrates is delivery with no backend, low latency and privacy by construction — which is what matters to clients whose data cannot leave a site or a jurisdiction.

Wishes by Hand on the App Store →

Applied modelling

Machine learning · 2024

Tailings facility risk benchmarking

A capstone project with a global mining-engineering consultancy, completed as part of a four-person University of British Columbia Master of Data Science team.

Roughly 300 recorded tailings storage facility failures and about 1,800 operating facilities, modelled on precipitation, seismic hazard, dam raise method and operational status. The deliverable was a reproducible pipeline and a prediction dashboard; the gradient boosting model reached an F2 of 0.826 on held-out test data.

The design choice worth noting is interpretability. The tool was built to pre-screen facilities for closer human review, not to hand down a verdict — which is the same reason it was usable at all.

How we engage

Fixed-scope build — a defined deliverable with written scope, acceptance criteria and a fixed price.
Ongoing maintenance — monthly retainer for content updates, dependency upgrades, monitoring and small changes.
Technical advisory — day or week rate: architecture review, AI agent workflow design, data pipeline design, interpretable risk modelling.

Request a quote