Mathew Stevens

25-package benchmark foundry; nine selected finalists with 52/54 required-deliverable failures across Codex and Claude trials

Model-discovered legacy UI workflows replay as typed capabilities with no model in the execution path

Terminal-Bench 3 AI eval with 6/6 clean reward-0 Claude Opus and Codex failures

AI Fund (Andrew Ng) build challenge, 1 of 12 engineers: B2B decision engine with deterministic margin, inventory, and policy guardrails

Crash/recovery conformance for agent runtimes and exactly-once external side effects

AI-agent external-action safety for retries, crashes, receipts, and exactly-once side effects

Defensive verifier-hardening audit for benchmark answer-key detector undercount

Eval-integrity audit for success-judge artifacts and prompt-injection ASR inflation

Prompt-injection containment by provenance, capability, and tool-argument role

CLI that detects model-alias behavior drift with frozen prompts and confirmation runs

Pure-Python pre-execution risk scoring and benchmark leakage audit for agent tool calls

Open pre-sign transaction safety for wallets and apps

Production-ready UE5 wallet adapter plugin

First Mobile Wallet Adapter SDK for Cocos Creator

Swift package for encrypted wallet deeplink signing

Native Android wallet bridge for Capacitor apps

Merged reconnect, SIWS, auth-cache, and sign/send PRs

Merged WebGL wallet auto-select UX PR