Computer engineering student at High Point University; engineering intern at Prebid.org, summer 2026. I build agentic systems — mostly tools that check other systems' work, and platforms people actually pay for.
Compiles a repository's own documented conventions into an enforceable ruleset and judges
pull-request diffs against it. Deterministic, keyless checks catch CI tampering; an LLM judge
handles the semantic rules, and never sees the PR title or description, so the author's framing
can't bias the verdict. Benchmarked at ~85% precision / ~39% recall against human reviewers on
held-out prebid/salesagent PRs, with recall landing within a point across three
independent test sets. On the PRs where they were active, the two general-purpose review bots
already running covered none of the human-flagged issues.
A Claude-driven buyer that runs a campaign end to end against a live AdCP sales agent: product discovery, budget planning behind a spend gate, media-buy creation — across hand-rolled MCP, A2A and REST transports. The model does the judgment; a deterministic core does the wire work, and the money commit sits behind a durable executor it never touches, with the wire idempotency key and the workflow id kept deliberately distinct so a seller outage stays retryable.
Driving it against the live reference implementation surfaced divergences that implementation's own test suite does not catch — three transports with three different unknown-field policies, and a one-line lookup-key bug that accounts for an open issue in the upstream tracker. 126 tests, none skipped. The protocol fuzzer it was built to feed is designed and not yet built.
Merges two colleagues' review tools into one pipeline — Chris Huie's review charter, detectors and pattern catalog as the rigor, Konstantin Mirin's scout → review → draft-only-post flow as the chassis. Sixteen review agents unioned into eleven canonical dimensions behind one synthesis and readiness gate, over ten deterministic detectors, with both upstream licenses and provenance tracked throughout.
Pointed at one of my own pull requests, it found a test that had silently stopped testing anything: an upstream signature change made the stub raise, the error was swallowed into a null, and the assertion it guarded never ran again. On its first full run it also caught a finding its own synthesis step had dropped — exactly the failure its ledger check exists to catch.
Three pull requests merged, thirteen open, and twenty issues filed against the protocol server itself — mostly correctness and security hardening rather than features: redacting webhook credentials from push-notification logs, sanitizing untyped error text before it reaches the buyer, returning the correct JSON-RPC code for an unknown A2A task, and a UTF-8 fix for a guard that crashed the suite on Windows. Most of these started as something one of the agents above turned up.
Built after Prebid's marketing team — then paying an outside agency ~$10K/month for WordPress edits — asked for an in-house alternative. Lets non-technical staff run website and mailing-list workflows in plain English, with a human approving anything that goes public. I wrote the onboarding guide and trained the team; it's now used by Prebid's entire marketing and engineering teams.
A commercial platform for professional Rainbow Six Siege teams, used by roughly half of pro-league
rosters: a binary .rec replay parser, a PyTorch computer-vision VOD analyzer,
AI-generated coaching reports, multi-tenant Firebase, a Discord scrim bot, Stripe billing, and a
desktop client. I run the engineering and the customer support.