TL;DR: My second DevFest Central Valley in Modesto was about who AI is for, from the mayor's opening to a keynote on AI as a community institution. In the buildathon I made Onceover, a first-look accessibility checker for small business websites, built in Google AI Studio with Gemini. It tells an owner what it found, where the rule comes from, and what it can't see.
Modesto Mayor Sue Zwahlen stood at the front of the room on Saturday morning and said something most officials wouldn't. She doesn't always know the how of technology. She knows the why.
Then she told a room full of developers, students and the plain curious: we need you.
That was the start of my second DevFest Central Valley, hosted by GDG Modesto with GDG Stockton, GDG Turlock and GDG Merced. GDG stands for Google Developer Groups, the local chapters that run these events. DevFest is their yearly community conference. The theme this year was Build. Scale. Secure. There were talks, labs, a scavenger hunt, and a short buildathon. I came for the buildathon. I'm writing about the room.
Know the why
The mayor's point stuck with me because of how she framed it. When people show up and share what they know with each other, tech grows in a community. That kind of collaboration can change things. She didn't need to understand a model to understand that.
I work in the California courts. Most of the people I build for don't know the how either, and they don't need to. They need the thing to work, and they need to trust it. Hearing a mayor say that out loud, in a room built for builders, set the tone for the day.
AI is an institution, so how do we take part in it?
Saphonia Foster's keynote was on leveraging AI for socio-economic justice. She described AI as an institution, like a library: something that holds and shares knowledge, and that a community gets to shape by how it shows up.
Her framework had three parts, and I wrote them down fast:
| Lens | Her question |
|---|---|
| Time | Are we spending our time with AI productively, or losing it there? |
| Voice | Are we using our voices to shape AI and AI policy? |
| Value | Are we using AI to lift up our communities, guided by our own perspectives? |
I kept running my own work through it. For a small business owner, time is the whole budget. Voice is whether the tools she uses explain themselves or talk down to her. Value is whether a tool leaves her better off or just more anxious. That last one ended up shaping what I built.
Test your output
The Building in the Present and Shaping the Future panel, from the Girls Who Code College Loops group at Stan State, gave me a short list I could have taped to my laptop:
- A security-first mindset
- Test your output
- Build accessibility in from the start
- Start a mentoring circle or community of your own I walked into the buildathon an hour later with that list still in my ears. "Test your output" turned out to be most of my afternoon.
Building for one small business owner
I write about my wife E. a lot, because she runs the kind of business most software forgets. She's a one-person wholesale baker, La Condesa Baking Company. Earlier this month I built her Cuenta, an invoicing and shelf-count app, after she lost (and then found) the notebook her whole business lived in.
This time the problem was a website. She wants one. She's also scared to make one, because she has read enough about ADA lawsuits aimed at small businesses that no website feels safer than a website she might get wrong.
That's a time, voice and value problem. She doesn't have the time to learn accessibility standards. Most accessibility tools speak developer, not baker. And a tool that tells her "you're compliant" might leave her worse off than no tool at all, because no automated check can promise that.
So in about three hours I built Onceover: a first look, not an audit. You paste in a page's HTML. A testing engine checks it. Each problem comes back as a card with the accessibility rule it breaks, a link to W3C's explanation of that rule, and a plain-language note on who it affects.
earlgreyhot1701D
/
onceover
An accessible first-look web diagnostic tool for small business owners. Combines 100% deterministic axe-core scanning in an isolated sandbox with prompt-isolated Gemini explanations. Delivers clear customer-impact guidance in plain language without compliance theater, legal claims, or raw code exposure. Evidence precedes conclusions.
Onceover
Transparent, deterministic accessibility diagnostics paired with small-business plain-language guidance—built with zero compliance theater.
Problem Statement
Small business owners—bakers, mechanics, florists, and boutique retailers—increasingly rely on their websites as the front door to their business. However, the web accessibility landscape presents them with two bad options:
-
Incomprehensible Developer Jargon: Existing tools like raw browser audits or enterprise compliance suites generate intimidating reports filled with CSS selectors, stack traces, and legal citations. Business owners without in-house engineering teams cannot understand or prioritize these findings.
-
Deceptive "Compliance Theater" Overlays: Predatory third-party plugins promise "one-line AI fixes" and guaranteed legal immunity. In reality, these overlays often disrupt screen readers, fail to fix underlying code flaws, and create false confidence while leaving businesses vulnerable to complaints and alienated customers.
-
The Guarantee Trap: No automated tool can certify an entire website as barrier-free. Yet, typical products either over-promise compliance or overwhelm users into inaction.
The
…The rules are written down in a short table, and they're enforced in code:
| It never says | It says instead |
|---|---|
| "You pass" or "compliant" | "We found nothing on the checks we ran." |
| A lawsuit risk score | The rule's level and who the problem affects |
| "Fix" | "Possible ways to address this (for human review)" |
| Nothing, when it can't check something | A permanent panel: "What Automated Checks Cannot See" |
A real testing engine, axe-core, decides what counts as a problem. Gemini only explains the finding, and it never sees the pasted page. If Gemini's explanation includes a banned word like "compliant" or "safe," the app throws it out and shows the citation alone.
The output needed testing. A lot.
The panel's advice paid off three times in one afternoon.
The first build scanned a copy of the page wrapped in its own template, which quietly added the language attribute my test page was missing. One of the six planted problems vanished before the scanner ever looked. Then the contrast problem went missing too, because the hidden frame doing the scan had no size, and the engine can't measure colors it can't lay out. It filed that one under "couldn't decide" instead of reporting it.
Both times the app looked fine. It just found fewer problems than were there. For a tool whose whole job is not giving false comfort, that's the worst way to fail.
The third one was funnier. When Gemini reviewed the code, it told me to swap the model from gemini-3.8-flash to an older one, and called the newer name "a hallucinated/invalid model string." Gemini 3.8 Flash is Google's current model. Gemini flagged its own successor as made up, with total confidence, inside a project about not stating what you can't source. The older model worked fine, so I left it. But I wrote it down.
Every line of code came from AI Studio. Claude is my build partner, and it planned with me and checked Gemini's ideas against the sources. Gemini handled the brainstorm, most of the code review, and the demo page with six problems planted on purpose.
The room after
At 1:45 we stopped building. Everyone in the hackathon room talked through what they'd made. It wasn't a formal demo, but I still had to stand up and explain mine. Afterward, one of the other builders came over and asked for the link.
That was the day in small. Someone builds a thing, someone else wants it, and now two people are talking.
What it can't do
- Tell whether alt text describes the picture correctly. It only checks that alt text is there.
- Read words inside an image, like a photographed menu.
- Test keyboard use.
- Show the checks the engine couldn't decide. That's next.
- Scan a live URL. You paste HTML.
- Tell you whether your business is covered by the ADA. That's a lawyer question. ## Try it
Onceover is live. Check the box, click "Load sample bakery page," and run a first look. Your pasted HTML stays in your browser. The only thing sent to the server is each finding's rule name, so Gemini can explain it. Nothing is saved. [TBD: how long the link stays up.]
I'll be back in Modesto in 2027. If you're anywhere near the Central Valley, come. Bring a laptop for the buildathon, but stay for the room.
Quick context if you are new here. I work in the California courts, running court operations for the county. I started building with AI in July 2025 and I have been learning in public ever since. I do not write the code. I direct, the agents generate, I validate and decide. I build the Clew Suite, a set of civic tech tools for making complex systems easier to inspect. That is the lens I am writing from.
Onceover is not legal advice and not an accessibility audit. It is a first look.
Built with AI assistance in October 2026. Talk notes are from my own notebook, so any errors in paraphrasing the speakers are mine.
AI Assisted. Human Approved. Powered by NLP.
Top comments (0)