Tony Nguyen — software engineer focused on deterministic evaluation for AI agents. Code is cheap now; verification is the scarce part. So every claim I ship maps to evidence you can re-run: a test name, a merged pull request, a live page. Nothing asserted, everything shown.
arcade-agent — an architectural guardrail for human & AI-generated code: 24 MCP tools plus a GitHub Action that reports architecture drift on every pull request. Listed on the official MCP Registry. Repo · arcade-agent.dev
Eval · Observe · Govern for agents — deterministic pass / warn / block gates for agent-written work, and a SigNoz-style observability stack that runs entirely on Cloudflare: OTLP ingest in a Rust Worker, AI traces with per-token cost. otel-worker
AI Kit LLC — done-for-you AI for local service businesses, bilingual EN / VI: receptionists that never miss a call, no-show reminders, review engines. Current pilot: RoofGuard Claims, document-sharing for roofing contractors and insurance claims. ai-kit.net · RoofGuard
Open source, weekly — contributing where maintainers can check the work: performance and concurrency fixes in Java and Rust ecosystems, plus one new browser mini-game every day. Also in development: HoloStage VR, spatial karaoke for Meta Quest.
Architecture recovery and smell detection for any codebase — tree-sitter parsing, quality metrics, and PR drift reports, exposed as 24 MCP tools and a GitHub Action for AI coding agents and humans alike.
github.com/tuannx/arcade-agent →OpenTelemetry ingest (traces, logs, metrics) in a Rust Worker, analytics on Iceberg / R2, with GenAI spans and token-cost tracking built in. No ClickHouse, no servers to babysit.
github.com/tuannx/otel-worker →An MCP server that gives AI agents control of the Servo browser engine — navigation, inspection and the accessibility tree. Think chrome-devtools-mcp, but for Servo. MIT.
github.com/tuannx/servo-devtools-mcp →Document-sharing between roofing contractors, their clients and insurance carriers — photos, scopes and claim paperwork in one place instead of scattered texts and email threads.
roofguard-claims.pages.dev →A local Spring Boot stack that rehearses AWS multi-region failover — Aurora Global semantics, JDBC failover, event ingestion — so the disaster drill happens on a laptop, not in production.
github.com/tuannx/spring-boot-multi-region-ha →One tiny HTML5 game, designed, built and published every day — vanilla Canvas, zero dependencies. Day 28 and counting. A shipping streak, kept honest by a public catalog.
Play the catalog →Stanford CoreNLP — eliminated an O(N²) quadratic copy in CoreLabel / ArrayCoreMap; merged upstream as PR #1530 after five review rounds (Oct 2026).
Apple ServiceTalk — root-caused a lost-demand race in PublisherBufferConcurrencyTest (issue #1672): stress runs showed 121 lost demands in 200k, zero in 1M with the fix. The maintainer invited an upstream PR.
arcade-agent in the wild — the first external adoption of the arcade-agent GitHub Action was merged by the maintainer of elasticsearch-analysis-vietnamese (PR #168).
Dataverse — a QuickPerf proof-of-concept that counts database queries in tests, catches N+1s and enforces SQL budgets; the approach was welcomed by the Dataverse team in upstream discussion.
Deep comparative reads from the engineering blogs of Netflix, Uber, Shopify, Cloudflare, Meta and more — each with my notes in Vietnamese and a link straight to the source.