Trevor Wattles

Trevor Wattles

Backend developer at J. Reuben Clark Law School and full-stack developer at EnrollPro. I build Node.js and Express APIs on MySQL and AWS, and I graduate from BYU in Computer Science in April 2027.

Vineyard, Utah · Open to full-time roles from May 2027

Selected work

4 projects

AI Study App

A study app built around students' own class notes. The AI fills out rough notes, turns them into quizzes and flashcards, and runs study sessions as Ace, a tutor that quizzes you instead of doing your work.

Developer · Jun – Sep 2026

JavaScript · PHP · CSS · OpenAI API · Playwright

Notable work

  • Built AI features that run through the notes: semantic search over notes using embeddings with a keyword fallback, vision support so the AI can read photos and hand drawings, and chat tools that let the AI edit a note or put it on the calendar with the event linked back to the note.
  • Made AI edits dependable: an edit the AI confirms in chat always lands in the note, later saves can’t overwrite it, and the change history stays bounded when the AI rewrites a large note. Each note can also carry its own AI chat, placed at the bottom, on the right, or turned off.
  • Built out the calendar: day and week views on one shared time grid, an agenda that says what repeats (“Repeats weekly on Mon, Thu”), recurring events with custom frequencies, and scheduling goals, reminders, and tasks straight from a note.
  • Reworked the canvas board, document view, and Explorer search, closed an HTML-injection hole in search snippets, and added 27 automated test files (22 unit, 5 Playwright end-to-end).

Private repo · Walkthrough available upon request

The Study home page with a prompt to turn notes into flashcards and quizzes, a row of quick-learn question cards, and the student's study spaces.
Study home, where notes become flashcards and quizzes
An Ace tutor session for a history course, where the AI introduces what the test covers and asks a multiple-choice question about Reconstruction, with a chat box to message Ace below.
Ace, the AI tutor, quizzing from a student's history notes

The hard part

Read aloud had to highlight the sentence being spoken, but the speech layer only reported progress as a fraction of each chunk of text, so the highlight ran ahead on short chunks, lagged on long ones, and couldn’t be turned back into a place in the note. I changed it to report a character offset, driven by the speech engine’s own word boundaries when it sends them and re-anchored to each chunk otherwise, so drift can never outlast one chunk. Then I mapped the note’s text nodes to that same string and back, which made click-to-jump possible. The highlight paints through the CSS Custom Highlight API, so it never touches the note’s markup and never ends up in the undo history or the saved note.

Home Price Prediction

Predicting Las Vegas home sale prices from 11,000 real MLS sales with an ensemble of machine learning models.

Winter 2024 · BYU CS 270

Python · pandas · scikit-learn · geopy · Google Geocoding API · matplotlib/seaborn · Google Colab

Notable work

  • Cleaned and encoded 11,000 Las Vegas home sales from the Multiple Listing Service (MLS), turning categorical fields, dates, and yes/no amenities into model-ready features.
  • Geocoded every address to latitude/longitude with a two-pass pipeline: free Nominatim lookups first, results outside Las Vegas filtered out, and the ~4,000 failures retried through the Google Geocoding API.
  • Modeled real estate “comps” with a location-only k-nearest-neighbors model (k=2 was best across 10 random splits), and used k-means clustering (k=5, chosen with silhouette scores) to group similar homes.
  • Combined a neural network (MLP), kNN, and k-means into an ensemble that cut error (RMSE) from $289,991 to $86,099, against a $325,883 baseline of guessing the average price.
Map of Las Vegas home sales colored by k-means cluster
Home sales colored by k-means cluster (k = 5)

The hard part

Our first neural network was barely better than guessing the average price. Sale prices ranged from $40K to over $15M, so a handful of luxury homes dominated the loss. We tried several transforms on the target before settling on trimming the 1st and 99th percentiles, log-scaling, and min-max normalizing, which gave the model a smooth distribution it could actually learn. The bigger gain came from location: once every home had coordinates, a nearest-neighbor model doing what an appraiser does, looking at what nearby homes sold for, outperformed using all the other features together. Ensembling it with the MLP and k-means clusters brought error down 70%.

JWT Pizza

The testing and DevOps side of a pizza-ordering web app, with automated tests, CI/CD pipelines that deploy to AWS, and live monitoring with alerts.

Testing, CI/CD, and AWS deployment · Sep – Dec 2025 · BYU CS 329, QA & DevOps

Playwright · Jest · GitHub Actions · Docker · AWS (S3, CloudFront, ECR, ECS) · Grafana Cloud · k6 · Node.js · MySQL

Notable work

  • Wrote 48 Playwright end-to-end tests for the frontend and Jest tests for the Express API, added user update, list and delete features test-first, and published coverage badges from CI.
  • Built GitHub Actions pipelines for both repos. Every push lints, tests and checks coverage. Then the frontend deploys to S3 behind CloudFront, and the backend ships as an ARM64 Docker image to ECR and ECS. The pipelines sign in to AWS with OIDC, so no access keys are stored.
  • Added Grafana Cloud metrics and logging to the service, covering requests by method, latency, auth successes and failures, pizzas sold, revenue, CPU and memory, with an alert on spikes in 5xx errors.
  • Load-tested login and ordering with k6, ran a chaos exercise and wrote the incident report, and pen-tested my deployment with a classmate.
Grafana Cloud logging dashboard for jwt-pizza-service, showing recent structured log events with auth headers redacted, a count of 72 exceptions, a pie chart of HTTP status codes, and factory request volume over six hours.
Grafana Cloud logging dashboard for the pizza service
k6 Cloud load test results showing 742 requests with no HTTP failures, a peak of 10.33 requests per second, a 150 ms 95th-percentile response time, and per-endpoint timings for login, ordering, and menu requests.
k6 load test of login and ordering, 742 requests
GitHub Actions run summary for the jwt-pizza-service CI pipeline, triggered by a push to main, showing a successful Build job (36 seconds) followed by a successful Deploy job (1 minute 11 seconds).
CI pipeline run, build then deploy to AWS

The hard part

The security test found two holes that together gave admin access. Any logged-in diner could call GET /api/user and get every user’s email and role, which showed who the admins were. And the password check only ran in the browser: deleting the password field from the login form let me sign in as an admin with just an email address. I added a role check so non-admins get a 403, removed a duplicate /api/user route that was overriding the real handler, moved password verification to the server, and hid admin links from other roles. Then I confirmed the fix with curl as a diner and as an admin.

Other projects

  • Neo4j Academic Papers — Loaded 100,000 arXiv paper records into Neo4j as papers, authors, and categories and answered questions about them in Cypher.
  • Anthem Gutters — The website for Anthem Gutters, the gutter-guard business I co-run in Utah.
  • Tweeter — A Twitter-style social app with a React frontend and a serverless TypeScript backend on AWS Lambda and DynamoDB, deployed with AWS SAM.
  • Chess Server — A multiplayer chess server and command-line client in Java, with a Spark HTTP API, live game moves over WebSockets, and MySQL storage.
  • Conference Search — Built with a teammate, it answers questions from 426 General Conference talks using keyword, semantic, and RAG search on Supabase, pgvector, and GPT-4o.
  • Gigglr — A random joke generator where users log in, save favorite jokes, and see other users' saves in real time over WebSockets, built with React, Express, and MongoDB and deployed on AWS.

Experience

  1. Apr 2026 – present

    Software Developer

    EnrollPro · Remote

    • Build full-stack features and testing suites for an enrollment and team-building CRM.
    • Diagnose and fix frontend and backend issues reported in customer tickets.
  2. Jan 2026 – present

    Backend Developer

    J. Reuben Clark Law School, BYU · Provo, UT

    • Develop and maintain the law school's information system.
    • Build and test REST APIs with Node.js and Express, manage MySQL databases, and work with AWS.
  3. Feb – Dec 2025

    Delivery Associate

    Amazon · American Fork, UT

    • Employee of the Month twice among 140+ employees.

Education

Brigham Young University

B.S. Computer Science, expected April 2027

GPA 3.93 (4.0 in major) · Dean's List

About

I grew up in Las Vegas and served two years as a volunteer representative in France, where I spoke French every day. Now I build software at two jobs: the backend of the J. Reuben Clark Law School’s information system at BYU, and full-stack features at EnrollPro. Building something new is my favorite part of the work, and lately that usually means AI. I also co-run Anthem Gutters, a gutter-guard business in Utah, and I built its website.