Community technologist/Haverstraw, New York
Daniel Aguilar, in a suit and tie, on an indigo background.

Structures decide. So I build better ones.

I build websites, tools, and public records. I have worked inside front-line services — on the staff of a Latino community nonprofit, helping people enroll in health coverage. Before I built for this work, I lived it.

Four projectsEvery one of them has a source you can open. Two languagesI shipped the Spanish translation rather than promising it. One federal docketI read it and checked every date and statute before launch. Both sidesI have worked inside a community nonprofit, and I have been a recipient of front-line services myself.

What I Build

Civic software /

I build sites that put documented, checkable information where people can reach it — filings, timelines, primary sources. Every claim traces back to something you can open yourself.

Community websites /

I build for nonprofits and small businesses. The site works on a phone first, it is findable in local search, and it is in more than one language when the community speaks more than one.

AI fluency /

I build with these tools in the open and publish the receipts: what they got right, what they got wrong, and how I checked. The discipline is checking what the machine hands you against evidence.

Built in the Open

On Tuesday, August 29, 2023—one week to the day before my first class in math and computer science at my community college—I typed my first prompt (a message to an AI) into ChatGPT. When I came back to the chatbot on October 1st, it was with a homework problem, and for the next three years the record looked like this: 874 active days, about 22 replies a day. Slow and steady. By three independent counts, 53 to 58 percent of every word I exchanged with ChatGPT was learning, and the biggest single slice was computer science and programming: Java homework and exam prep.

AI replies per month Conversational AIAgentic AI
June 2026: 13,305 agentic AI replies ↓
5,000
10,000
2023
2024
2025
2026
Aug 2023 – Apr 2026: about 22 replies per active day From May 2026: 345 per active day
Each bar is one month of replies from ChatGPT (conversational) or Claude (agentic), counted from my own exported records; July 2026 runs through the 19th. The chart counts replies, not tokens (the units AI companies bill by): a big token count proves exposure, not value. The full accounting, with its method and independent audits: AI With Receipts →
The numbers behind the chart
MonthConversational AI (ChatGPT)Agentic AI (Claude)
Aug 20234—
Sep 20230—
Oct 2023205—
Nov 2023308—
Dec 2023573—
Jan 2024459—
Feb 2024297—
Mar 2024402—
Apr 2024668—
May 2024396—
Jun 202465—
Jul 2024776—
Aug 2024287—
Sep 2024341—
Oct 2024564—
Nov 2024596—
Dec 2024993—
Jan 2025905—
Feb 2025669—
Mar 2025672—
Apr 20251,502—
May 20253,064—
Jun 2025459—
Jul 2025333—
Aug 2025370—
Sep 2025612—
Oct 2025500—
Nov 2025503—
Dec 2025446—
Jan 2026213—
Feb 2026539—
Mar 2026905—
Apr 2026395—
May 20265203,093
Jun 20264713,305
Jul 2026415,694

In May 2026, after three years of learning, the curve broke. I started using Claude, Anthropic’s AI, and went from conversational AI to agentic AI: instead of asking questions and getting answers, I was directing teams of AI agents that research, build, and draft whole projects—websites, apps, and, above all, civic and legal advocacy—which I reviewed. The pace went from 22 replies per active day to 345. The single biggest thing I have ever built with AI is the Bad Faith Prosecution wiki—a public, sourced, plain-language record of a criminal prosecution.

The Record

2026Live

I designed and built this public record.

Bad Faith Prosecution

A public record of the Broadview Six case: independent legal research and citizen journalism. Every claim traces back to the filing it came from. I read the federal docket and checked each date and statute before any of it shipped.

badfaithprosecution.wiki →

2026Live

I built this site, and the writing on it is mine.

Grown From Concrete

My own writing, including an honest accounting of my AI use — the full corpus exported, classified by an analyzer I built, independently audited, then verified again from scratch on a competitor’s tools.

grownfromconcrete.org →

2024Archived

I designed and built this site for a client.

Serie 32 Grocery

A Dominican-owned bodega. I built the local search around how neighbors actually look for a corner store and wired in DoorDash ordering rather than a checkout the store did not need.

Read the full entry →

2017Archived

I ran this website as a member of the staff.

Latinos en Axión STL

I ran the website and built out the Spanish translation — and worked Cover Missouri outreach, helping people enroll in health coverage under the Affordable Care Act.

Read the full entry →

Most people who build for community organizations have never worked inside one. I have. I have also been the person those organizations exist for. I know what it is like to need the website to do one specific thing for somebody who is counting on it, today, with no budget and no time. And I know what a bad form costs when you are the one who needs it to go through.

That is also why I will not hand you something you cannot check. If I say a page is faster, there is a number. If I say a tool got something right, there is a source. What I learned as a grassroots organizer is shaping what I build next: a tool that lets a small group run a real meeting inside a Signal chat — discussion, debate, motions, and a recorded vote.

Start a conversation

Leave a Reply