Building practical, interesting tools — and proving they work.
Training operations & reporting · AI-assisted development · Testing & verification
I find where people actually get stuck, build the smallest thing that fixes it, and prove it works before anyone else has to.
- 31free resources, all downloadable
- 23browser tools, each tested before it shipped
- 1live AI product, built and deployed alone
Start from the problem
I don't build training to check a box. I find where learners actually get stuck, then design the smallest resource that fixes it.
Make it usable
Job aids, classroom tools, and trackers a facilitator can open and run immediately — clean, low-friction, easy to trust.
Measure and tighten
I track readiness, scores, and follow-up so training keeps getting sharper instead of staying static.
What’s here
About me
The long way around, and what it taught me.
I came into this the long way around. I started out teaching martial arts — up to a dozen classes a day, every age and every skill level — and that is where I learned the thing that still runs my work: people don’t learn from what you tell them, they learn from what they practice. Retail, mobile sales and call centres followed, and in every one of those rooms I ended up being the person who trained everyone else.
The first thing I do with any process is look for how to make it better — but I won’t propose a single change until I understand it completely. Learn it inside and out, then improve it. That order is why the things I build tend to hold up.
What I do now
At NYU Langone Health I design and run new-hire and upskilling programmes across Hospital Billing, Patient Access, Accounts Receivable and Customer Service — and I build the machinery underneath them. The Excel, SharePoint and Power Automate trackers the training team runs on are mine: scheduling, attendance, assessment scores, learner readiness, programme outcomes. I’ve trained and coached more than 4,000 people, and revamped roughly 60% of the programme I inherited.
Then I started building software
I built my own AI assistant, Zareli, and took it to a live deployment. Its first specialist reads spreadsheets properly — it tells you what is structurally broken, why, and what to do about it. I own the architecture, the product decisions and the verification; AI coding tools write the implementation under my direction.
The same habit runs through everything here: 23 single-file tools, each driven through a headless browser and checked against assertions before it shipped. I don’t publish what I haven’t tested.
The rule I build against
One idea sits under all of it: competency, not dependency. Most AI is designed to make you need it more. I wanted the opposite — something that hands the work back to you a little more capable than it found you.
That isn’t really about software. It is the same job as standing in a classroom: you are finished when people don’t need you any more. A trainer who makes themselves indispensable has failed at the actual assignment. It is why every tool here has an Edit button instead of a support request.
The part people ask about
I don’t have a college degree. I’d rather say that plainly than let it sit in a gap on a resume. What I have instead is a body of work you can open right now, a product in production with real users, and seven years of getting people competent at things they found hard. I’m working toward DP-900 and PL-300 because I’d rather add the credential than argue about it.
Let’s talk
Everything in one place — the resume, how to reach me, and the documents behind the work.
- Name
- Tyler Locke
- Location
- Boynton Beach, FL
- Working style
- Remote-first · open to hybrid with relocation
Open to roles where the job is finding an operational problem, building something small that fixes it, and proving it works. If that’s adjacent to what you’re hiring for, it’s worth a conversation.
Documents
The written record behind the work.
How I approach learning design
Solve the real problem. Keep tools simple enough that another facilitator can pick them up and trust them. Design for recall, not coverage. And measure readiness — not just attendance.
How I approach AI
AI accelerates drafting, organizing, and prototyping. A human verifies accuracy against the source of truth before anything reaches a learner — and a human stays accountable for every claim the output makes.
Everything in the free tools library, and the product on the Zareli tab, came out of the same working method. It is not complicated, but it is the part people ask about most.
Working with it, not having it work for you
Most people use AI as a vending machine: describe the thing, take what comes out. That works when you already know exactly what you want and the answer is easy to check. It fails badly when the thinking is the actual work — because a fluent answer to the wrong question looks identical to a fluent answer to the right one.
The shift is asking it to interview you before it answers, to argue with your reasoning, and to name the weakest part of what it just produced. That turns it from something that writes for you into something that thinks alongside you — and the output stops being something you have to take on trust.
What I’ve learned
- Confidence isn’t accuracy. A wrong answer arrives with exactly the same certainty as a right one. Nothing in the output tells you which you got.
- Deterministic beats clever. If a thing can be calculated rather than predicted, calculate it. A parser can’t be confidently wrong.
- Direct the build, own the judgment. The architecture and the standards are mine. The typing is not the valuable part.
- The goal is needing it less. If I finish a session no more capable than I started, something went wrong.
Where this actually led
Working this way is what made a real product possible. Zareli is an AI assistant I architected, built and deployed on my own — the clearest evidence I have that the method produces something that works, not just something that sounds good.
Selected work
A short list, on purpose — the projects that best show how I work. Each opens and runs in a browser, and each has a case study behind it. Everything else — all 31 of them — is in .
No tools match that search. and try again.
How these are builtNothing ships until it passesShow the standard
Each tool is tested in a headless browser before it goes near this portfolio — not opened and eyeballed, but driven through its actual behavior and checked against assertions.
- Behavioral tests, not screenshots. A scripted browser clicks through the real interaction path — start a round, run a timer, submit an answer, save, reload — and asserts the result at each step.
- Zero console errors is a gate, not a goal. Any error, warning, or failed request fails the build. Several tools were rewritten rather than shipped with a warning in them.
- Tested at real widths. Every layout is swept from 360px to 1920px and checked for horizontal overflow, wrapped controls, and clipped content — which is how I found a tab strip that broke in an 80-pixel window nobody would have thought to check.
- Offline is verified, not assumed. Every file is scanned for external requests. A tool that quietly loaded a web font was caught and rebuilt with a local stack.
- Accessibility measured, not assumed. Contrast ratios computed against WCAG AA on every visible text element — compositing transparent backgrounds, because reading them as white gives the wrong answer. Every interactive element has a visible keyboard focus ring, and every tool honors
prefers-reduced-motion. Verified across all twenty tools, not sampled.
The same habit runs my AI work: I don’t ship what I haven’t verified, and I’d rather find the failure myself than have someone else find it.
Handing these to another trainer? The maintenance walkthrough explains how to keep them updated.
Tools you can run today.
Nearly all of these open in a browser and keep working offline — no install, no account, nothing needed once you have the file, and every one has an Edit button so you change the questions to yours. One is live and needs wifi; it says so below, and it names the offline tool that stands in for it.
Not finding it? — that’s how I decide what to build next.
Zareli.ai — an AI that reads your spreadsheet instead of guessing at it
Most AI is built to make you need it more.
Zareli is built to make you need it less.
An assistant that knows when to stop guessing and call in a specialist. The first specialist reads spreadsheets — and really reads them, so that what it finds in your file teaches you something you keep. Built and deployed independently.
A sample workbook, opened cold
Nothing was typed. This is what Zareli had already worked out by the time the file finished loading.
Sample workbook built for demonstration — fabricated data, real defects, real output.
What it does
Open a workbook and Zareli reports what’s structurally broken — a formatting rule that quietly stopped applying, a formula reading a blanked merged cell, a reference that drifted out of range — naming the sheet and the range, before you ask it anything. Then it explains the finding in plain language, and turns it into something you keep.
The decision that defines it
Most tools point a language model at a spreadsheet and ask it to work out what’s going on. That’s expensive, it’s slow, and it fails invisibly — a model that misreads a formula answers just as confidently as one that reads it correctly.
Zareli’s audit is a parser and a graph walk. No model call. Select a stage to see where the boundary sits.
What building it took
Deterministic where it can be
Reproducibility mattered more than cleverness. A parser can’t be confidently wrong, costs nothing per run, and can be tested. The model stays where its strengths actually are — explaining.
I own the judgment; AI writes the code
Architecture, product decisions, verification, and the standards governing what the system may claim are mine. Implementation is written by AI coding tools under my direction.
It only says what it can prove
If a defect can’t be demonstrated from the file itself, it isn’t reported. Written down before there were users, in a founding document I still hold it to.
Built, deployed, and maintained alone
FastAPI, React/TypeScript, Postgres, Anthropic models, on a VPS behind Cloudflare with GitHub Actions CI/CD. No team, no investors.
Invite-only, on purpose
Around fifteen people use Zareli today. I’m funding it myself, and I’d rather it work properly for fifteen than badly for fifteen hundred. Access is by request while it’s small.
The full product — what it does, how it works, and how to request access — lives at zareli.ai. What’s here is the part that matters for hiring: a clear problem, a deterministic answer wherever one is possible, and a human accountable for every claim it makes.