AzimuthFBO · Kevin Deveau

Projects

Skills shown Swift · SwiftUI · macOS · Software Development · Large Language Models (LLM) · Prompt Engineering · LaTeX · Product Design

Not a résumé generator. A native macOS app that runs your job search: it finds roles in your field and area, scores how well each one fits you, tailors your résumé to the posting from your real experience, drafts cover letters for you to approve, and learns from your edits so its first drafts keep getting better.

What I did

Built a coach that learns from every revision: it reads past corrections and proposes updates to the writing guidance, with evidence, so each new draft starts closer to what the user wants
Built lead generation that finds open roles matching the user's field, level and location, and scores each with a 1 to 5 fit rating against their real experience
Tailored every résumé to the job description only from recorded facts, flagging any requirement the user's experience does not cover
Drafted cover letters in the user's own voice that nothing sends without their approval
Produced polished, one-page PDFs through a LaTeX builder, then checked each for page count, readable text and section headings
Built the whole app solo in about 46,800 lines of Swift, with about 11,000 lines of tests

Highlights

Gets better the more you edit it

Every revision a user asks for marks something the first draft got wrong. The app collects those corrections, turns repeated ones into private writing rules, and applies them to future drafts, so the user stops giving the same note twice.

Scores how well each job fits before you apply

Each role is compared against the user's experience, eligibility and location, and gets a fit rating with its strengths and gaps written out. The user spends time on the roles worth applying to.

Swift · SwiftUI · PDFKit · Codex CLI · Claude Code

Skills shown React · TypeScript · PostgreSQL · Supabase · Stripe · Payment Processing · Full-Stack Development · Database Design

The platform a local home-services company ran its business on: customers book and pay online, and staff manage jobs, memberships and customers in one place. I led the software end to end, from the database to production.

What I did

Took the platform to production for 100+ users, with online booking, recurring memberships and Stripe payments
Integrated Stripe for online card payments, refunds and memberships, processing thousands of dollars in real customer revenue
Made double-booking impossible: a slot is held during checkout and locked in the database, so two customers can never get the same time
Built an admin suite of about two dozen views, with a full customer record and a customer-health score that drives retention emails
Recovered the 30 to 40% of ad conversions that browser tracking lost by also reporting them from the server
Shipped through two automated test gates with 100+ end-to-end and 250+ unit tests

Highlights

Ran in production for a local home-services company: 100+ users booking and paying for real.

Scheduling that fits the real length of every job

Each service has its own duration, which changes with the size of the vehicle, plus a buffer between jobs. Customers only see times that actually fit, and when two people try to book the same slot at once, only one gets it. Skills: System Design · Database Design · PostgreSQL

Blocks a release if the database update fails

A new version of the site only goes live after its database changes succeed. If they fail, the previous version stays up instead of breaking.

React · TypeScript · Supabase · PostgreSQL · Stripe · Vercel

Skills shown C++ · Compiler Design · LLVM · MLIR · ANTLR · Software Testing · CI/CD · Git

A full compiler for Gazprea, a language first designed at IBM in Markham, Ontario, with vectors, matrices, intervals, tuples and streams, and values immutable by default. Built by a team of four, it ranked first in the class, and the team was invited to an online knowledge exchange with HSBI in Germany to present it. It takes source code to native programs through MLIR and LLVM, keeping the program's meaning exactly intact at every step.

What I did

Built nine ordered checking passes between parsing and code generation, from scope and type resolution to cast validation, each reporting typed errors with line numbers
Modelled eight kinds of types, including vectors, matrices, intervals, tuples and structs, with promotion, casting and conversion as separate services
Generated code through six standard MLIR dialects and lowered it to LLVM IR in seven stages, with a dump after each for debugging
Wrote a C runtime for arrays and vectors: each is a length and a data pointer on the heap, grown as needed and freed automatically when its scope ends
Optimized the compiler's back end with LLVM's optimization pipeline, shrinking generated code by 85 to 90% across the test suite
Ran the team of four like a product team: features split into a dependency graph so work ran in parallel, every pull request reviewed, and a protected main branch gated by CI

Highlights

Ranked #1 in the class; the team was invited to an online knowledge exchange with HSBI in Germany to present the compiler.

Cut generated code by up to 90% with LLVM optimization

The compiler produced correct but bloated code. Wiring LLVM's -O3 pass pipeline into the back end took about 30 lines and shrank the output by 85 to 90% across every test, including the hardest stress tests. It showed the value of building on a mature compiler toolchain instead of optimizing by hand.

Designed the compiler as ordered passes, each with one job

Each pass can assume the ones before it succeeded: types are resolved before they are checked, and names are resolved before qualifiers are enforced. Adding a new rule has an obvious place, and a failure points at the exact stage.

Kept a growing compiler correct with 2,101 end-to-end tests

Every feature shipped with tests that compile a real program, run it and check its output, all under a memory checker in CI. The suite grew to 2,101 programs, including deliberately hard stress tests, and AI fuzzing found edge cases the hand-written tests missed. Skills: Software Testing · CI/CD

Ran a four-person compiler build like a small engineering team

We met weekly in person for a stand-up and co-working session, run like an office: review progress, re-plan, and assign the next tasks. The work started from a thin end-to-end skeleton, then split into features mapped by what each depended on, so we could plan the critical path and keep all four people building at once. Every change was reviewed, formatting and linting ran on every commit, and nothing reached main without passing CI. Skills: Project Management · Agile Methodologies · Team Leadership · Code Review · CI/CD

C++ · ANTLR4 · MLIR · LLVM

Skills shown Compiler Design · C++ · MLIR · LLVM · ANTLR · GPU Programming · Computer Architecture

A learning project to understand how ML compilers like Triton and XLA work inside, built alongside AI coding tools. One kernel has to become very different code for a CPU and a GPU without being written twice.

What I did

Designed the kernel language and the compiler's architecture, keeping how a kernel is scheduled (tiling, unrolling, fusing) separate from the math
Directed AI coding tools through the implementation, reviewing each stage against tests I specified
Took the CPU path all the way to x86-64 AVX2 assembly on disk
Took the GPU path to NVIDIA tensor-core instructions inside the compiler; nothing is executed or benchmarked yet
Built one custom MLIR layer that both the CPU and GPU paths translate from, plus a type checker that reports errors at the exact line

Highlights

One compiler core for both CPU and GPU

The CPU and GPU need very different code, but both start from the same internal layer. Adding the GPU target meant new translation rules, not a second compiler.

C++17 · ANTLR4 · MLIR · LLVM · TableGen

Skills shown Assembly Language · RISC-V · Computer Architecture · Low-Level Programming · Debugging · Problem Solving

Snake with its core game engine written in raw RISC-V assembly. It runs in RARS, the RISC-V assembler and runtime simulator, through its Keyboard and Display MMIO Simulator: interrupt-driven, with no operating system or libraries underneath.

What I did

Wrote the core game engine in 1,114 lines of assembly: the game loop, movement, collision detection, apple spawning, scoring, a countdown timer and three difficulty levels
Drove the game entirely from hardware interrupts: key presses and timer ticks arrive through one handler that saves and restores every register
Read the keyboard and drew to the display through memory-mapped I/O, redrawing only the three cells that change each move

Highlights

Handles hardware interrupts safely, by hand

With no operating system, the program installs its own interrupt handler. It frees a register by swapping it with a reserved save slot, saves every register the game uses, works out which device fired, and restores everything before returning.

Talks to devices directly through memory-mapped I/O

The keyboard and display are memory addresses. The game polls their ready bits, positions the cursor and writes characters itself, and only redraws what changed because every display write waits on the device.

Wrote a complete game engine with no operating system or libraries

The game loop, movement, growth, collision detection, scoring and timing are all written by hand in assembly, on top of nothing but the processor and its devices.

RISC-V assembly · RARS

Skills shown Product Management · Agile Methodologies · Scrum · Stakeholder Management · Requirements Analysis · Python · FastAPI · Next.js

A multilingual reading app built for SFU Linguistics that lets K-12 students read the same book in the language they read best, with comprehension checks and progress metrics for teachers. It supports 200+ languages, including Pashto, Somali and Amharic, which classroom software rarely covers.

What I did

Led a team of seven as product owner, shipping 300+ user stories over five two-week sprints
Made books readable in 200+ languages: a teacher uploads a PDF, and text recognition, a teacher check and translation do the rest
Built comprehension checks: questions are generated from the book's text, students answer in the app, and attempts are gated on reading progress
Gave teachers and cohort leaders metrics on student progress, comparing each class and student over time
Cut translation start-up from 20 seconds to near zero by loading the model once and reusing it

Highlights

Delivered to SFU Linguistics as the research contribution itself: 300+ user stories shipped across five sprints by a team of seven.

Checks that students understood what they read

After a chapter, questions are generated from the text at three difficulty levels. Students answer in the app, and the server decides when a quiz is available, so a refresh cannot get around it.

A teacher checks scanned text before it is translated

Pages where text recognition is unsure wait for a teacher to fix them, so mistakes are not translated into every language.

Next.js · TypeScript · FastAPI · Python · PostgreSQL · Meta NLLB

Skills

My strongest skills are teamwork, communication and collaboration, and working closely with people is my favourite part of the job. I do my best work on fast-paced teams. I take ownership of features end to end, from talking with users to running them in production, and I'm happy to take on whatever the team needs that week.

Core Languages9

Web Technologies19

Systems & Compilers13

Databases & Cloud11

Data & Analysis6

Mathematics7

ML & Data Science12

DevOps & Tools15

Testing & Quality8

Process & Collaboration16

Design4

Support & Operations7

Aviation Club Management System

Founding Software Engineer · 2025

Keeps flying clubs running every day.

Software for flying clubs, flight schools and FBOs.

Visit azimuthfbo.ca

What it is

AzimuthFBO is operations software for flying clubs, flight schools and FBOs. Scheduling, dispatch, maintenance, training, membership and billing live on one record, so nothing is entered twice.

What I did

Architected, built and scaled AzimuthFBO from zero to production: one multi-tenant platform for scheduling, dispatch, maintenance, training, membership and billing, serving organizations of up to 1,500 members Skills: System Design · Full-Stack Development · Next.js
Won unanimous board approval over competing products, pitching with live demos, video demos and marketing material I produced Skills: Sales · Presentation Skills
Built the lead discovery pipeline feeding AzimuthFBO's sales and marketing outreach, sourcing and qualifying flying clubs, flight schools and FBOs across North America Skills: Lead Generation · Data Analysis
Migrated 49,000+ records from legacy systems, validated by reconciliation reports Skills: Data Migration · SQL
Led client onboarding end to end: discovery, data mapping, rehearsals, parallel running and go-live Skills: Client Relations · Project Management
Lead the product vision and roadmap, translating operator needs into scoped, scheduled and prioritized releases Skills: Product Management · Requirements Analysis
Designed the UI/UX across the product and marketing site, from Figma designs to a reusable component system Skills: UI/UX Design · Figma
Built the CI/CD pipeline and release process: every merge redeploys a staging environment, production is reached only by promoting staging, and database migrations deploy through CI Skills: CI/CD · DevOps · GitHub Actions · Release Management
Designed the PostgreSQL schema and its indexing, with each organization's data isolated in both the application and the database Skills: Database Design · PostgreSQL
Wrote the SOPs and runbooks behind data migration, client onboarding, sales demos, disaster recovery and production operations Skills: Technical Documentation
Lead the AI initiative at AzimuthFBO, streamlining support, development and operations Skills: Artificial Intelligence (AI) · Team Leadership

HighlightsAn AI agent that fixes reported bugs without database accessMembers report problems from any screen. An AI agent picks up the report, shows its plan as it works, and attaches the code change it proposes. Because it reads text from the public, it never gets database access; it works only through a narrow, locked-down interface. Skills: Artificial Intelligence (AI) · Application Security
Checks every aircraft is loaded safely before it fliesPilots enter fuel, passengers and bags and see the aircraft's weight and balance against its certified limits as they type. The same calculation runs again on the server before anything is saved, so the number the pilot saw is the number on record. Skills: TypeScript · Software Testing
Tracks flight training from first lesson to checkrideInstructors build syllabi of ordered lessons and record each student's progress, grades, stage checks and checkrides. Students see where they stand in their course, and instructors see their whole roster. Skills: Database Design
Ships to production safely, every timePull requests run only the tests for the parts of the code they touch, and anything the selector cannot place runs the full suite. Promotion to production runs every test, and a nightly run keeps it green. Database migrations plan before they apply, cap how many apply at once, and stop if an environment's credentials are missing. A disaster recovery plan sets recovery time and data-loss targets for each kind of data, with point-in-time restore and a runbook for each failure. Skills: Software Testing · Disaster Recovery · Production Support
Built the migration that moves flying clubs and schools off legacy systemsEvery rule about the old data is checked against all of it before any import runs. That is how hidden cancellations (37% of one history) were found and how phantom flights were avoided. Every import is safe to re-run: a second full run changes nothing. Skills: Data Migration · Quality Assurance