Portfolio

Real systems. Specific claims. Nothing that can't be checked.

Everything below is something that exists and runs, not a mockup. Where I have a number, I'm giving you the number.

Tier 1 · Flagship engineering

Systems where being wrong is not an option

RF engineering platform

Wavefront AI

A headless engine that turns a natural-language request into a constraint set, places radios against it, computes terrain-aware path profiles and coverage rasters with real diffraction, grades the result against NFPA or IFC compliance rules, and renders a report with full provenance. An external agent can call it over MCP and gets back only figures the engine itself derived and verified.

the actual hook RF values aren't plain numbers. They're a branded type, RfFigure, that only the engine package is allowed to mint. A dependency-free CI script (check-boundaries.ts) fails the build if any other package reaches for that function, and it's verified by deliberately planting a violation and watching it get caught. assertEngineDerived() re-hashes the inputs at every MCP and report boundary. A language model producing a fabricated coverage number literally cannot make it typecheck.

The engine also withholds what it can't substantiate: without terrain data it reports a warning instead of claiming line-of-sight, and every synthesized antenna pattern is disclosed on any result that uses one. A newer module (@wavefront/graph) models the conducted half of the same link budget, VSWR, insertion loss, distance-to-fault, so field diagnostics can run offline on a phone with no server and no propagation engine at all.

TypeScript, 13 packages 339 tests, 65 suites, 0 failing MCP tool surface DSATUR channel assignment PostgreSQL + PostGIS
Open-source AI governance

Project Starfish

A portable, deny-by-default governance overlay that drops onto any existing Claude or AI agent build: every tool call, agent action, and capability passes a single policy decision point before it's allowed to run. It governs the file system (boundary-confined reads and writes), shell execution (catastrophic commands denied by rule), network and MCP access (admitted but tainted, no exfiltration), and secrets. It can govern Claude Code itself, through its own hooks.

704 tests green Published on npm Apache-2.0 Electron + React desktop UI Model-agnostic
projectstarfish.ca →
Student productivity platform

Fazoozle

"A Second You. For school." Fazoozle turns a student's syllabi into one daily instruction: open it in the morning and it tells you what to do today, in order, with how long each thing takes. An 11-stage ingestion pipeline extracts every deadline, weight, grading rule, and undated obligation from raw course documents (including nested weight tables and month-grid schedules), and a planner rebuilds the term-long schedule backwards whenever a day slips.

the actual constraint Two of the source syllabi explicitly forbid uploading their contents to third-party services, which made data residency a first-class design constraint, not an afterthought. The whole stack consolidates onto Azure Canada Central: Static Web Apps for the PWA, Container Apps for the planner and the Python ingestion service, Durable Functions for the multi-stage extraction pipeline, and Postgres with row-level security for tenancy, one subscription and one identity model instead of a five-vendor patchwork.

Beyond the core planner, it includes a Study Workspace with a shared, growable knowledge base (concepts, mastery tracking, cards, quizzes), NeuroTrace 3D, and the professor, a verified conversational tutor built for the actual student it's designed around.

NeuroTrace 3D is an interactive neuroanatomy pathway simulator: pick an event, a pinprick on the right foot, blood pressure rising, and watch the signal traced through a rotatable, zoomable, explodable 3D body, receptor to nerve to plexus to spinal cord to cortex, with every reflex and output that branches off on the way. The core rule is that no pathway is allowed to teleport: every connector between origin and destination has to be on the route, even the ones a textbook summary skips, and it's enforced as a build gate, not a style guide. A topology checker validates every route against the 3D model itself before it ships. The first run of that checker caught 86 joins where the animated signal crossed undrawn empty space, in one case a 49cm jump from pituitary to kidney, and each was fixed with the real, named structure the signal actually travels through.

the actual hook "The professor" is a tutor the student can type or talk to that is not allowed to teach her something wrong. Every explanation runs through a five-stage pipeline: retrieve the source passages, build a verified step skeleton, generate the explanation with every sentence tagged to a citation, then run a separate verification pass at temperature zero that labels each claim supported, partial, unsupported, or contradicted against its own cited source, and rewrites or cuts anything that doesn't hold up before she ever sees it. It answers only from her certified course material, the OpenStax anatomy & physiology textbook, and NeuroTrace; if the answer isn't in her materials, it says so instead of guessing.

The part that evolves with her is the learner model underneath it: Bayesian knowledge tracing per concept, an open model she can see and dispute, and a strategy bandit (Thompson sampling across analogy, worked example, Socratic questioning, visual/3D, and more) that tracks which teaching approach actually works for her, on this topic, based on whether she still remembers it days later, not just whether she nodded along in the moment. A forecast model flags which upcoming lecture material she's statistically likely to struggle with and runs a short "heads-up" session beforehand. It also knows the difference between a real problem and a bad day: told something meant to provoke a safety response, it has an explicit, tested rule for defusing that stays a person, not a script, while still routing genuine distress to a real crisis line.

Azure Static Web Apps + Container Apps 11-stage Durable Functions ingestion pipeline Postgres with row-level security PWA with iOS web push 3D neuroanatomy simulator, build-gated against its own topology Verified RAG tutor: retrieve → generate → verify → repair Bayesian knowledge tracing + Thompson-sampled teaching strategy
Tier 2 · Production, under Pinnacle Tech Projects

Shipped for clients who can't afford it breaking

Pinnacle Tech Projects is the Canadian AI foundry I run out of Barrie. These are real, live builds, not case-study fiction.

Construction · government

CCAIP

Automated compliance checking for building-permit applications. AI extracts and classifies the submission; a deterministic code engine issues the verdict, built to the NRC standard, with a human keeping final authority.

Document intelligenceHuman-in-the-loop
Telecom · field ops

FieldCapture AI

Turns WhatsApp field-crew messages and photos into structured daily reports and Monday.com records with no workflow change for the techs. Its own classifier was found defaulting 84% of messages into one bucket by inspecting a real 31-message transcript; fixing the default alone dropped it to 26%.

Workflow automationMonday.com
Regulated · professional services

Arranger

An offline-first operations platform for Ontario funeral homes: compliance gates written as declarative rules, government Form 15 mapping, payments and accounting export, bilingual EN/FR from the ground up.

Offline-firstCompliance rules

See the full body of work at Pinnacle Tech Projects →

Tier 3 · AI that checks its own work

Research pipelines that catch their own mistakes

SR&ED technical claims

IntelicoreAI SR&ED research pipeline

A self-correcting AI research workflow for CRA SR&ED tax claims, built for IntelicoreAI's grant advisory work. A fact-check pass caught seven real errors in its own first draft against primary CRA sources, including a wrongly assumed repealed grind and the wrong refundability percentage. While reviewing a client's RF site files as evidence, a date-walk across their own records surfaced two live regulatory violations nobody had flagged: unlicensed operation inside a DFS-mandatory band, and a transmitter set to an invalid channel frequency.

CRA-facing researchSelf-correcting pipeline
AI visibility & SEO tooling

GetAIVisibility & CheckAIListing

Tooling that measures how often AI answer engines (ChatGPT, Perplexity, Google AI) actually name a business, built on Supabase and Cloudflare Workers with a real free-to-paid funnel. One of my own sites, FlowNorth.ai, has already been named unprompted by ChatGPT in a live buyer-intent query, the exact effect this tooling is built to measure and improve.

Supabase + CloudflareDataForSEO
Tier 4 · The AI automation properties

A working laboratory, not just client work

A network of live AI-automation and AI-consulting sites, each one both a real offer and a live test of what makes a site actually get found and cited.

FlowNorth.ai

Toronto-facing AI automation agency: n8n workflows, AI voice agents, RAG knowledge systems. Named unprompted by ChatGPT in a live web-search answer for "AI automation consulting firms in Canada."

flownorth.ai →

IntelicoreAI

Grant advisory and AI implementation for Canadian businesses: SR&ED, IRAP, and provincial funding programs stacked against AI adoption projects.

intelicoreai.ca →

AICapacity & AIImplemented

AI strategy, implementation, and workflow automation for Canadian small businesses, two distinct brands testing two different positioning angles on the same underlying capability.

aicapacity.ca →
Tier 5 · Range

Not everything is enterprise software

Prysio

A large-scale content and workflow generation system spanning 19 industry niches, 16 buyer personas, and 7,047 distinct workflows, built to test how far structured AI content generation can scale before quality breaks down.

Journeys Start

A fantasy fiction world and serialized book project, built and published as its own site. Proof that the same discipline for structure and internal consistency applies whether the system is an RF engine or a fictional universe.

Next

Want to talk about a build like one of these?