Can AI solve homelessness? No. But the question went somewhere.

To be clear, I'm not a developer or a data scientist. I'm a paralegal. I work in legal aid, I got curious, and I write these notes mostly to figure out what I learned. This one started with a question I'm a little embarrassed by and got more useful as it shrank.

01The oversized prompt

It started as a thought experiment in a NotebookLM notebook: could you use AI to solve homelessness in Connecticut? I knew the answer was no. I asked it anyway, to see where aiming at something that big would actually take me.

It didn't take long to fall apart. "Solve homelessness" isn't one problem. It's a few hundred problems sharing a word. Every time I tried to make the prompt specific, it broke into something smaller and more ordinary.

02What it collapsed into: 211

The big question kept landing on the same quiet answer. Connecticut already has a way in for someone in crisis. You dial 211, the statewide line that routes people to housing, mental health, and crisis services. There are Coordinated Access Networks for homelessness. There are dozens of agencies already doing the work.

So the real problem isn't that we need AI to solve this. It's that the help is scattered, and a person in crisis usually has to already know it exists to find it. Once I saw it that way, the project got doable. I stopped trying to solve homelessness and started trying to see the safety net that's already here.

03So I mapped it

The map at the top is what I built. It plots 33 services across the Hartford–New Haven corridor, pulled from 211ct.org, sorted into three categories: mental health, autism services, and post-eviction or housing help. You can filter, search, and click any point for the agency, its hours, and a phone number. The map and the table stay in sync.

A few things I picked up building it:

• Geocoding is most of the job. Turning 32 street addresses into coordinates, and caching them so I wasn't asking for the same one twice, took longer than anything else. It's also where I learned the most.
• You don't always need a map library. I plotted the points onto a plain SVG by hand instead of pulling in something heavier. At this size it was enough, and I understood every line of it.
• Picking categories is a decision. Choosing those three buckets, and deciding what counts as "post-eviction help," shapes what the map says. The structure isn't neutral.

04I called the real eviction API

This is the part where I actually wrote code against real data. My research kept pointing at Connecticut's open data portal and claiming eviction filings were public, sortable by neighborhood, even usable as an early warning. I didn't want to take that on faith, so I made the call myself.

The dataset is real. It's called "Connecticut Weekly Evictions by Census Tract," it lives on data.ct.gov, and it's compiled by Princeton's Eviction Lab with the CT Fair Housing Center. One short request to its API came back with clean data: eviction filing counts by census tract, week by week. There's something satisfying about learning that public-interest data like this is just sitting there, one request away.

Making the call also corrected something the summaries had skipped. The data is historical, not live. The fields are built around a 2020 pandemic baseline, and the copy I pulled was last refreshed in December 2025. That's not the real-time, updates-every-Monday feed my notes described. It's a research dataset that looks backward. The "catch the spike and knock on doors before the court date" version isn't what this endpoint is.

05A thread I pulled: the representation gap

One category on the map, post-eviction help, sent me down a side path that's worth its own note. Connecticut's Right to Counsel program gives some tenants a free lawyer in eviction cases. I wanted to know how well it worked. The data told two stories at once.

Before the program, about 7% of tenants had a lawyer in eviction court. In the original program zip codes, that went up to roughly 18%. More than double, and a real gain. Landlords, meanwhile, have a lawyer about 80% of the time. So the program works, and most tenants still face court alone. Both are true.

The part that taught me the most had nothing to do with charts. My research notes had some sharp, specific numbers in them: exact shares of eligible tenants left unserved, multipliers for "informal" evictions that never reach a courtroom. I went to trace them back to the original evaluations and couldn't. So I set a rule. If I can't tie a number to a primary source, it doesn't go on the chart. The wider point, that most eligible tenants still go unrepresented, I make in words instead of a number I can't stand behind.

06What the AI made — and why I left it out

Somewhere in the middle of this I let AI tools turn my research into a set of infographics. Full-color posters, big headlines, clean icons, confident percentages. They looked like something an advocacy group would put out.

That was the problem. The numbers on them didn't agree with each other, or with the primary sources. One poster claimed an 85% success rate, another 67%. Some homelessness figures I couldn't trace at all. The polish didn't make them accurate. It just made them look accurate. A clean headline over a number nobody checked is worse than a plain one, because it borrows trust it didn't earn.

So none of those images are on this page. I kept the one chart I can defend and I'm describing the rest in words. Useful for thinking, not ready to publish. That's the distinction I most want to hold onto from this.

07What I'd want to learn next

• A Spanish version. A lot of the people most likely to need 211 in this state speak Spanish. English-only is the wrong default. I'm bilingual, so there's no excuse not to fix it.
• Live data instead of a snapshot. Hours and phone numbers go out of date. Keeping something like this current is a skill I don't have yet.
• A plain-language way in. Most people can't name their problem in service-category terms. "I might lose my apartment" should find the housing resources without anyone needing to know the jargon.

I'm putting this up because the failed version of the question turned out to be the useful part. It pointed me at what's already here. If you work in this world and I got something wrong, tell me.