Ryan Russon

An automation tool that reads a SOC 1 Type 2 report and produces the firm’s filled-in review workpapers, flagging uncertainties for review. A SOC 1 review can include up to 2 hours of manual reading and scanning; the tool removes much of the manual searching, with a reviewer still verifying the result.

Uses an algorithmic, deterministic matching approach (no LLM layer), and accuracy is measured against existing completed reports. In testing, the tool exhibits zero high-confidence match errors. The suite continues in 02.

The other two tools in the suite, built on the same architecture as the SOC 1 pipeline.

The CPE tracker turns a folder of continuing-education certificates — different providers and layouts, some scanned — into the firm’s filled compliance spreadsheet, flagging anything uncertain for review.

The proposal generator builds a complete engagement package — cover letter, marketing body, engagement letter, merged PDF — from internal templates and data sheets. Both were 10–20 minute manual processes, now made near-instant.

A history trivia quiz app — React frontend, Node/Express service, MongoDB Atlas, live scoring over WebSocket. Runs on EC2 under pm2 behind Caddy for TLS.

CI/CD pipeline: GitHub Actions runs frontend and service test suites in parallel; both must pass before a deploy job that activates on a push to main. GitHub OIDC authenticates with IAM role assumption — removing the need for AWS keys in repository secrets — and the build produces a tarball and pushes it to S3, triggering deployment through AWS SSM. If a deploy fails, the server automatically rolls back to the previous release.

Frontend developer on a 6-person team maintaining two production applications for BYU Law School — the school’s public website and an internal data-management app — serving 500+ users. Migrated both from React to Svelte using AI-assisted tooling, keeping feature parity throughout, and maintained on AWS (ECS, S3, Route 53, CloudWatch) in coordination with backend developers.

Built user lookup and input validation into the internal app to cut manual data entry, and enhanced search capabilities and site layouts among other updates to the public site — linked below.

Two agents solving Wordle from a shared belief state — a Bayes filter that prunes the 2,315 possible answers after every guess. One picks guesses risk-neutrally, maximizing expected information gain; the other is risk-averse, minimizing the worst case. Benchmarked over every possible answer, and against a human player.