AI developer tooling, Test architecture, Reliable LLM features
I build software across backend, frontend and testing, with a growing focus on how people and AI agents work together. I’m interested in new interfaces that make that collaboration useful and controllable, with the context, clear expectations and verification needed to get reliable work done.
Looking for Agent and developer tooling, test and evaluation infrastructure, applied AI engineering, or senior full-stack work on AI-native teams. Open to conversations
Perspective · AI tooling: Context and clear expectations for useful agent work. Relevant capabilities and proof are marked. Change perspective
Three featured examples with the problem, my part and the key decision, then more write-ups on BenLive, my personal site. Each is labeled as current-role work, a personal project or first-person writing.
The C# adapter I architected in my current role to run Selenium or Playwright under the same test logic: the interface contract, backward-compatible release gates, parity tests, and where assistants are and are not trusted.
Problem
A large VB.NET Selenium suite gates releases, and rewriting it in one go would pause regression coverage.
My part
I architected the C# migration onto an adapter that runs Selenium or Playwright behind one interface.
Key decision
The language move and the driver move stay separate: tests target the adapter, the driver is configuration, and a ported test passes on the original driver before it runs on the new one.
Status
Port under way. The code is private; the sketch and the write-up show the architecture.
Test logic (C#)
Adapter interface
driver chosen by configuration
Selenium driverPlaywright driver
Illustrative sketch of the architecture in the public write-up; not employer code.
Current role · Test architecture + AI workflow + Context & team rules
Fallback chain, circuit breaker, layer-specific retries, cost allowlist and a privacy-preserving rate limiter, each traced to a real failure on my own site.
Problem
Free-tier LLM providers fail in ordinary ways: 502s, sudden rate limits, models retired without notice.
My part
I built the resilience layer behind my site’s public chat, in one Firebase Function.
Key decision
An ordered model chain with a per-model circuit breaker, so a visitor gets an answer from a smaller model before an error from a bigger one; only allowlisted models run.
Status
Live on my personal site. The backend code is not public.
Personal project · AI product reliability + Full-stack delivery + Test architecture
Personal tools for recruiter chat, prompt refinement, problem solving and agent configuration, designed around clear expectations and reviewable results.
Problem
Make useful AI tools that people can direct and understand, across more than one isolated experiment.
My part
I design and maintain my personal AI Lab with AI coding assistants: Chat About Ben, Promptpad, Problem Solving and Agent Agenda.
Key decision
Keep expectations and prompt contracts explicit, expose results for review, and give each tool controls suited to its job.
Status
Public personal projects on BenLive. Not employer work.
Personal project · AI workflow + AI product reliability + Interface & spatial
A personal agent-first OS I’m building: agent identities, memory, files and flow sequencing live on disk as readable JSON and Markdown, with inference dispatched through an observable queue.
Personal project · AI workflow + Context & team rules + Full-stack delivery
I designed the serverless backend, accessible vanilla-JS frontend, cross-browser tests, media pipeline and gated deploy, and I review and maintain them; AI coding assistants carried out the plans I approved.
Personal project · Full-stack delivery + Test architecture + Interface & spatial
Choose the perspective closest to your role, and the career record below highlights the parts it draws on.
AI TOOLING
Context and clear expectations for useful agent work.
At work, I set up Kiro and own its supporting tooling for an eight-person automated regression testing team, following the same work with Amazon Q Developer. Personally, I work with dots, Claude Duet and Muse alongside Claude Code and Codex, and build AI Lab tools for chat, prompt refinement and agent configuration.
Start with the AI Lab and assistant-rules evidence above; the record below separates QA-team tooling ownership from my personal work.
Personal AI projects are covered in the proof and capabilities sections.
QA Automation Engineer III
March 2022–present
AI-assisted development-tooling point person for an eight-person automated regression testing team: set up Kiro, the team’s current AI tool, and own its supporting tooling, as before for Amazon Q Developer, whose version-controlled Rules, Profiles and Prompts I authored.
Architected the C# migration of a legacy VB.NET test suite onto a cross-driver adapter that runs Selenium or Playwright under the same test logic; the port is under way.
Built and shipped an internal documentation and onboarding app that consolidates the team’s scattered docs into Markdown, with a web view built from the codebase.
Drive weekly regression triage end to end, converting manual test paths into automated tests.
Independent Web Developer · McNulty Web Apps
June 2018–February 2022 · 2012–2014 (separate period)
Carried responsive full-stack client projects from the first conversation through launch and support, translating ambiguous requests into practical frontend, backend and deployment decisions.
Frontend Engineer · Menuat
February 2016–May 2018
Shipped digital menu systems from concept to cloud-hosted production, translating brand requirements into responsive, accessible experiences on low-performance, TV-connected hardware.
Covered project management in the interim, owning scoping, delivery and support across the client base until a permanent PM was hired.
Full-Stack Developer · Enix Networks
September 2012–August 2013
PHP page servers, contact forms and data handling for client sites.
Seven disciplines, each with practices, tools and evidence you can open.
01 AI workflow
02 Context & team rules
03 Test architecture
04 AI product reliability
05 Full-stack delivery
06 Interface & spatial
07 Collaboration
01 / AI workflow
AI-assisted engineering workflow
I work with agents by giving them relevant skills, context and clear expectations, then checking what they produce. My current personal work spans dots, Claude Duet and Muse; Claude Code and Codex are my primary direct coding interfaces. Earlier projects include Local Crew, a local-model orchestrator.
Give agents the relevant project context, skills and working agreements
Choose the interface and model for the work, with a clear handoff between agents
Make requests, constraints and expected results explicit before a run
Let agents carry work forward within agreed boundaries
Keep changes reviewable and check the result against the request
Record what changed, what was verified and what remains open
Useful for teams that
are adopting AI coding assistants and want reviewable diffs, explicit authority and verified handoffs rather than unbounded agent output.
At work I set up Kiro, the QA team’s current AI tool, and own its supporting tooling, as I did for Amazon Q Developer before it; that and the AI-assisted porting in the driver migration are QA-team scope. GitHub Copilot is my personal VS Code experience, going back to its beta. Everything else here is personal practice on my own projects and public site.
02 / Context & team rules
Context engineering and team AI rules
I turn scattered information into context people and agents can use. For an eight-person automated regression testing team, I set up Kiro and own its supporting tooling, as I did for Amazon Q Developer. I codify internal engineering standards and procedures in agent-usable documentation, and built a documentation app that brings team material into Markdown and a web view.
Rules, Profiles and Prompts reviewed in pull requests like code
A regression an assistant introduces becomes a new rule
One documentation path for new engineers and new assistant sessions
One canonical policy file that each assistant’s instruction file points to
Docs change in the same pass as the behavior they describe
Agent-facing guides such as llms.txt generated from one manifest
Useful for teams that
keep knowledge in people’s heads or scattered docs, and want new engineers and new assistant sessions to start from the same reviewed context.
The Amazon Q Developer Rules, Profiles and Prompts, the Kiro tooling and the internal documentation app are QA-team work for my current employer. The canonical CLAUDE.md and AGENTS.md pattern, Claude Code skills and the llms.txt manifest come from my personal projects.
03 / Test architecture
Verification and test architecture
I start with a question: how could this passing result be false? At work, I architected a legacy suite’s move to C# on an adapter that runs the same tests through Selenium or Playwright. The port is under way. In my personal AI Lab, I audit test runners so interrupted batches cannot report success, and turn bugs I find into automated checks.
Test runners audited so a killed batch cannot report green
Every hand-found bug becomes an automated check
Test logic never knows which browser driver runs underneath it
Ported tests keep running on the original driver until a switch is judged safe
Page objects expose intent, not locators and waits
Weekly regression triage, driven end to end
Useful for teams that
carry a legacy suite, are changing drivers or frameworks, or are letting AI assistants touch code, and need a passing build to mean something.
The adapter migration and regression triage belong to my current QA Automation Engineer III role. Runner audits, cross-browser suites and the bug-to-check rule come from my personal AI Lab.
04 / AI product reliability
Reliable, honest LLM features
I build AI features with clear expectations, visible results and room for the person to direct the work. In my personal AI Lab, chat replies name the model that answered; versioned prompt contracts keep behavior explicit; and an allowlisted model chain with a circuit breaker is designed to return an answer from a smaller model rather than fail.
Clear expectations and versioned prompt contracts that make agent behavior reviewable
Chat replies labeled with the model that actually answered
Allowlisted models, a fallback chain and a per-model circuit breaker
System prompts kept server-side behind versioned prompt keys
Separate server and client retry policies, each traced to a real failure
Interfaces that help people guide the work, inspect results and build trust
Useful for teams that
put language models in front of users, where cost control, honest attribution, privacy boundaries and graceful degradation have to be designed in from the start.
Personal AI Lab work on my own public site.
05 / Full-stack delivery
Full-stack product delivery
I take web products from the first request through deployment and support, in TypeScript, JavaScript, React, Next.js, Node.js, C# and PHP. As an independent developer I carried client projects from the first conversation to support; earlier I built PHP client sites at Enix Networks. My recent personal work includes a serverless AI backend that I designed, built with AI coding assistants and maintain as its only engineer.
Serverless backends on Firebase Functions and Firestore
Rate limiting keyed on an HMAC of the visitor’s IP, never the raw address
Personal-site deploys gated by model probes and post-deploy smoke checks
One shared module where a configuration had been forked
Client delivery from first conversation through deployment and support
PHP page servers, contact forms and data handling
Useful for teams that
need one engineer to carry a feature across backend, frontend, tests and deploy, and to explain the tradeoffs between those layers.
The Firebase and serverless work is my personal AI Lab. Independent Web Developer, June 2018–February 2022. Full-Stack Developer at Enix Networks, September 2012–August 2013.
06 / Interface & spatial
Interface, motion and spatial UI
I build responsive, accessible interfaces, from advanced CSS and design systems to spatial UI. At Menuat I shipped animated digital menu systems on low-performance, TV-connected hardware. For my own site I directed a Codex-assisted rebuild of a Three.js and WebXR gallery, with hand-tracking grabs, intent-safe teleport and an AR miniature, and kept a route that needs no 3D.
Animated digital menu systems designed for constrained, TV-connected hardware
Brand requirements translated into responsive on-screen experiences
Hand-tracking grabs, intent-safe teleport and a hand-held AR view
Design systems built on oklch, clamp() type and container queries
Contrast checked in both themes; focus kept keyboard-safe
A stable, readable reduced-motion version of every effect
Useful for teams that
build visual or immersive products that must stay readable, accessible and smooth on real devices, not only in a desktop browser.
Frontend Engineer at Menuat, February 2016–May 2018. The WebXR gallery and site visuals are personal, AI-assisted projects that I directed and verified.
07 / Collaboration
Communication and handoffs
I connect the people and disciplines a project needs: users, design, frontend, backend and testing. At Menuat I was the frontend intermediary between users, design and backend. I make expectations clear, write down decisions and leave handoffs someone else can pick up; those habits also shape how I collaborate with agents.
Alongside the driver migration, I codify the team’s internal engineering standards and procedures in documentation agents can use. The versioned Rules and Profiles make the adapter contract and expected checks explicit; people own the architecture and release decisions.
The QA team’s repository-owned AI rules require every behavior change to arrive with its matching test update, and changes to those rules are reviewed in pull requests.
Each model call waits for a click, so end-to-end tests can mock the model and assert the control state at every step. Those tests check the controls; the model’s variable answers are a separate question.
The rule I use to pick models, the cheapest one that clears the bar for the task, with a fallback, reproduced locally before shipping, is the rule the public chat runs on.
Across my AI Lab’s delivery surface, assistants carry context and execute plans I have approved. I decide what to build, set the constraints, review the plan and the diff, and own the outcome.
Before each release of my personal site, the deploy path probes every live model with a tiny request; afterward a smoke script checks status codes and headers.
AI Lab apps send a prompt key, never a prompt, so system prompts are not shipped to the browser and a visitor can choose only from allowed keys. That narrows what a visitor controls; it is not a guarantee against prompt injection.
In the brief for my site’s launch reel I asked for implementation jargon to be replaced with plain language that tells the viewer what they can do, then reviewed the on-screen copy Claude rewrote.
The principles behind the work, each argued in the write-ups linked above.
Intersectional and cross-discipline experience
01
Ask how a green result could be false
Why: Ask how a green result could be false
A runner that treats a killed process as a pass looks green right up until it matters, so every passing result has to be honest about how it could be wrong.
02
Harness before model
Why: Harness before model
I choose the working loop first, judging its context, change control, verification and handoff, and only then debate which model runs inside it.
03
Context decides the result
Why: Context decides the result
Model quality sets a ceiling. The context an assistant starts with decides how close the work gets to it.
04
Write the team’s AI rules down
Why: Write the team’s AI rules down
Every team using assistants already has rules. They belong in a reviewed file, not in each engineer’s habits.
05
Spend inference only on request
Why: Spend inference only on request
My AI Lab apps do not spend a generation the user did not ask for; the person deciding what happens next is the interface, not a limitation on the AI.
06
Degraded, not down
Why: Degraded, not down
A visitor should get an answer from a smaller model before they get an error from a bigger one.
07
A stale document is a bug
Why: A stale document is a bug
Assistants execute against documentation the way they execute against code, so docs change in the same pass as the behavior they describe.
Teams where this combination fits
Engineering teams formalizing how they work with AI assistants
Assistant rules, prompts and review practice versioned in the repository, with harness and model choices made per task. For an eight-person automated regression testing team I am the AI-assisted development-tooling point person and authored its Rules, Profiles and Prompts.
Context & team rules + AI workflow + Collaboration
Product teams shipping LLM features to real users
Allowlisted models, fallbacks, honest attribution and privacy boundaries have to be designed together with a tested full-stack implementation.
AI product reliability + Full-stack delivery + Test architecture
Test-infrastructure teams modernizing a legacy suite
Frameworks or drivers are changing under existing coverage, release gates have to hold during the change, and assistants need written rules before they touch tests.
Test architecture + Context & team rules + AI workflow
Product teams that need a senior full-stack engineer
One engineer carries a feature across backend, frontend, tests and deploy, and can explain the tradeoffs between those layers.
Full-stack delivery + Interface & spatial + Test architecture
07 / Next conversation Let’s talk about the work.
Looking for agent and developer tooling, test and evaluation infrastructure, applied AI engineering, or senior full-stack work on AI-native teams. Florida, USA · Remote · Open to conversations.