Applied AI engineer who ships agents with their evals attached, now building a prototype WhatsApp support agent at LaunchGood. Before that I spent a year at a commercial equipment lender, taking a pricing system through a security-scoped IT intake to corporate handoff and co-designing a risk scorecard that collections managers work from daily, while running a multi-tenant payments platform a real business depends on. My McMaster MEng project is an AI study tool for neurodivergent learners whose verification gate withholds any claim it cannot cite.
Joined as a senior engineer in the pod that owns AI initiatives at a global crowdfunding platform. My interview submission was a deployed prototype with its own evals, Zakat-Eligibility Triage.
Building its WhatsApp support agent, now a prototype. It answers campaign creators’ questions in their WhatsApp groups from the support documentation, keeps staff approval for any action that touches platform data, and was re-planned around leadership’s feedback on the day of the first demo.
Designing it as a TypeScript service on the Anthropic SDK with structured outputs, evals and prompts versioned in git from the first commit. Group messages are treated as untrusted input, payout details are read-only to it, and every answer category has its own kill switch.
Built one bilingual TypeScript application to replace 13 Excel pricing calculators across 14 manufacturers and 9 pricing engines; rate configs are versioned Postgres rows with an audit log, so repricing needs no deploy.
Carried it through enterprise SDLC: a nine-document security-scoped IT intake, a leadership review and an IT pre-review, then a handoff of source, calculation spec and 16 written rulings to corporate IT, which owns it and is deploying it from that documentation.
Validated its finance core (payment, rate and IRR functions) against the booking system of record, reproducing 45 funded deals to within half a cent given each deal’s booked rate and structure.
That validation caught wrong terms in 7 of 9 pricing engines, inherited from the legacy Excel calculators the business had been quoting from. I corrected them the same day and pinned the fixes in a 1,370-test suite with mutation-tested assertions.
Did this while carrying a vendor-sales equipment-finance book across 10+ vendors: over $4M funded in each of April and May 2026 against a $2M monthly target, and over quota every month from month two.
Built Incite, a white-label multi-tenant booking-and-payments platform on a 74-model Prisma schema and 196 REST API routes: scheduling, waitlists, Square payments, memberships, ticketing and per-tenant theming.
One launch client runs its entire operation on it in production: $45K+ CAD processed since launch across 640 registered users, 1,380 confirmed bookings and 105 active memberships, with no parallel manual system.
Money is the part that cannot be wrong: idempotent payments, webhook signature verification and tenant isolation enforced in a tenant-scoped data-access layer, guarded by a build-failing test for unscoped models.
Every branch-protected pull request must pass 8 required CI checks: 16,000+ Vitest cases, two Playwright suites, a Square-sandbox payment run, Gitleaks secret scanning, and a replay of every migration from an empty database that refuses destructive DDL.
Rebuilt my training app, SKomp Forge, into an agentic coach across 11 Swift 6 modules, shipping to TestFlight: 36 tool schemas pinned in both Swift and Python and diff-tested for drift, every write a typed proposal the lifter confirms by tap, and full-duplex voice holding P90 313ms text-to-first-audio under a 1,500ms budget.
Junior Web Developer
May 2021 - Jun 2022
Giftcash Inc.·Remote
As one of three engineers in a 30 to 40 person development organization, migrated a legacy Django monolith to Node.js on AWS Lambda for per-request billing, tuned PostgreSQL indexes to kill the sequential scans behind balance lookup, and stood up Jenkins and GitHub Actions CI/CD.
PythonNode.jsAWS LambdaPostgreSQLCI/CD
With two colleagues, designed a collections risk scorecard on six years of delinquency history, selecting the model by customer-grouped cross-validation. Its five bands run from 0.647 to 0.017 average default probability on the scored live book, a 38x spread with no reversals.
The SVP announced it org-wide in Sept. 2026 after it re-scored a 1,748-contract book and ranked 1,023 contracts. Collections managers work from its daily list, and the scorer reproduces all 1,023 rankings in plain Excel, so the business can audit any call without the code.
TypeScriptNext.jsPostgreSQLPythonMutation testing
Mentored 100+ learners through algorithms, systems design and interview preparation, running every engagement end to end since 2024: acquisition, pricing, curriculum design, live sessions and follow-up review.
TDDpytestVitestPlaywrightMutation testingTLA+ model checkingZ3/SMT2
// education
Master of Engineering (MEng), Computing & Software
Sept 2025 - 2027 (expected)
McMaster University · Hamilton, ON
Mac Study Companion, an AI study tool for neurodivergent learners, co-supervised by Dr. Richard Paige and Dr. William Farmer and accepted by both supervisors in Sept. 2026. It turns lectures, readings and the course calendar into cited notes and spaced review cards, plus a Socratic tutor that asks questions and never hands over the answer.
Designed first for students with ADHD, following a 25-item evidence-graded learning checklist: minutes-long capped sessions, one card on screen, deliberately delayed correctness feedback (Feddermann et al., 2018), no lost streaks or penalties, and each deadline reduced to one next action.
The research contribution is a fail-closed verification gate, since a hallucinated note is worse than none. A Z3/SMT encoder checks completeness, source support, confidentiality and a recall floor; the release protocol is model-checked in TLA+ (4 properties, 2,647 states), and a broken mutant spec must fail in CI.
On a 10-lecture corpus the gate raised citation precision from 0.604 to 0.960, with recall and leak rate held by construction. Runs as 9 hexagonal microservices on Redis Streams, 2,489 tests.
A+Simple Type Theory
AMicroservice Architectures
Bachelor of Applied Science (BASc), Honours Computer Science
