Anyone can build an election results tracker now. Ask your AI assistant for one and you’ll have a page that scrapes the county sites before your coffee gets cold. We wanted to build the one that answered the question everybody in New York’s 17th District was asking and nobody was measuring.

NY-17 was one of the most closely watched House races in the country, and the most-watched in the state: Republican Mike Lawler against Democrat Cait Conley. And everyone covering it knew that one community could swing it. Rockland County’s Heimish neighborhoods vote in huge numbers and almost entirely one way, yet from the outside they’re nearly invisible. Lawler was counting on exactly that. “I have over 30,000 Orthodox Jews coming out to vote,” he told Punchbowl News a week before the election. “They can’t capture that in the poll.” One public poll of the race was taken entirely over Sukkos, when much of that community wasn’t picking up the phone.

So we built one. lawlervsconley.com had a live election-night count of NY-17, with Rockland’s Heimish districts pulled out and watched on their own, against their own history. That’s hyper-local, and it takes a lot more than scraping county websites.

Both campaigns were watching it.

lawlervsconley.com● live · 1:11 AM
Lawler concedes
According to ABC News, Lawler conceded to Conley. The race has not been called; votes are still being counted.
Conley (D)
133,219
Lawler (R)
121,881
Rockland
Westchester
Putnam
Dutchess
Behind the margin Rockland, net votes
Heimish districtsLawler +18,307
Districts with Heimish pocketsLawler +4,326
Rest of RocklandConley +11,963

Finding the districts

Before the night could mean anything, we had to know which of Rockland’s 307 election districts to watch. There’s no public list. A district’s lines don’t follow neighborhoods, and the census doesn’t ask which shul you attend.

We built the groups from several data signals: data from the Board of Elections, obtained through FOIL requests; open-source community-level and geographic data; our own knowledge of the community; and AI help in research, estimation and modeling. Every district ended up with one of three labels: Heimish districts, districts with Heimish pockets, and the rest of Rockland.

research — grouping Rockland's districtsinternal
Board of Elections data
Open community-level data
Open geographic data
Our own domain knowledge
AI research, estimation and modeling
one label per district · 307 districts
Heimish districts 46Districts with Heimish pockets 24Rest of Rockland 237
✓ checkedinternal model, against 2016–2024 results

An internal model, using the groupings we developed, was then used to reproduce the past: we ran it against Lawler’s 2022 and 2024 margins, the midterm turnout dips in 2018 and 2022, and where his 2024 gains really came from. The results confirmed our groupings, and turned up the number that framed the whole race: roughly 12,000 of Lawler’s 14,600-vote improvement from 2022 to 2024 came from the Heimish districts. We kept the model internal and we deliberately did not post a forecast or needle, but every estimate on the site was sanity-checked against it.

lawlervsconley.com — map307 districts
Stony Point 14
Rest of Rockland
L 166C 102
Heimish districts
Heimish pockets

Agents in every role

We’re an AI shop, so we wanted as much of the site as possible to run on AI, autonomously.

Several agents ran it, each in its own role. The data feeds were automatic; one agent watched the official results sites and feeds like AP’s to sanity-check our numbers (more on that below). Others ran the Live Updates feed, some on their own and some with a human editor. Not one post was typed by hand.

One rule decided which: if a post changed what readers believe about who’s winning, or relied on a source the system hadn’t pre-vetted, a human had to confirm before it was published. Everything else the agents did alone. So posts came about in three ways:

  • Fully autonomous, about 70% of election night’s posts. Two writer agents posted on a schedule from the days before the election through election night, every 6 to 14 minutes while results were moving.
    • Context, armed with a rich fact book, read the race through the lens of the news and the politics.
    • The Count, armed with historical and live data, read it statistically.
    • Every draft had to pass automated checks. Each number had to trace back to the live data, the fact book or an exact quote from a source the writer had pulled up that round, and anything hinting at who lives where was blocked. A draft that failed twice never posted.
    • When the data itself looked wrong, the writers didn’t post about it. They sent the orchestrator agent a private alert.
  • Agent-proposed, human-approved. The orchestrator, the agent in charge of running the site and talking to the team, monitored constantly. When something looked off, like a county quiet for too long, AP showing votes we didn’t have, or a writer’s alert, it messaged the team with a proposal, and a human editor said yes or no.
  • Human-asked, agent-written. An editor messaged the orchestrator with an idea; it did the research and the writing, and the editor approved the text. Asked late on election night whether Heimish turnout was keeping up, it found the night’s headline: the Heimish districts had passed their 2024 presidential-year vote, in a midterm.
lawlervsconley.com — Live Updates● live
  1. The count

    New Square's count already tops its entire 2024 vote, at 116%, while Kaser's stands at just 11% even though every Kaser district shows as reported. That gap leaves the Heimish districts at 49% of their 2024 vote so far, against 72% in the rest of Rockland.

    Source: rockland county resultsWritten by the Count · no human in the loop
  2. Context

    Lawler's campaign said this morning it has gone to court to have Westchester's emergency ballots counted in "one central location," where the campaign, election monitors and lawyers can watch. No court filing has been reported yet. The county's Board of Elections has said every ballot will be counted but that delays may occur, Gothamist reported.

    Source: x.com, gothamist.comOrchestrator proposed · editor approved
  3. The count

    The Heimish districts have now passed their 2024 presidential-year vote: 19,344 votes, more than 2024's 19,302, in a midterm. That's 63% of their active voters, against 57% in the rest of Rockland. In 2024 the pattern ran the other way: 63% in the Heimish districts, 70% elsewhere.

    Source: enhancedvoting.comEditor asked · agent wrote · editor approved

The guardrails were partly the models’ own and partly ours, down to a quiet mode that slowed the writers once the count stopped moving. With both campaigns reading, the line was simple: report the count, give it context, never predict.

No admin dashboard

The usual way to run a site like this on election night is a back office: a dashboard, a login, someone with a hand on every lever. We didn’t build one.

The orchestrator agent ran the website. It monitored every county feed against the live site, tailed the server logs, deployed fixes, and checked AP and Decision Desk HQ for a race call whenever the count changed. It had a clear decision process for what it could do on its own and what it had to bring to us. Fixing a feed and shipping the code it did on its own, about 30 times over the night. Putting a number from anywhere but the county on the site, or anything that touched the outcome, required a human’s permission.

When Lawler conceded before any desk had called the race, something we hadn’t anticipated, the orchestrator asked whether it could build and deploy a new banner to announce the concession. The banner was live about six minutes after ABC News first reported it.

orchestrator — election night
|
|
ask editor
Enter Westchester from AP until the county feed takes over?
%
editor
yes
%
|
|
WestchesterAP fallback

A live site you can’t rehearse

Here’s the problem with a live results site: you can’t test it ahead of time. The feeds don’t exist until the polls close. Counties change formats and addresses without telling anyone. Whatever is going to go wrong goes wrong at 9:38 PM, in front of everyone.

So we planned for it like risk, not like a bug list. There was a written playbook for the night, mapping every symptom to a next step. There were fallbacks: if a county’s own feed failed, its numbers could come from AP, clearly labeled as AP, until the county’s feed came back. And the orchestrator was sanity-checking the site against AP and the counties in real time, ready to take mitigating action within minutes.

This approach held up pretty well in real time. Putnam published its results under a different election ID from the other counties, and Westchester moved to a new results site this year. Each time, the orchestrator caught the gap and filled it from AP until it fixed the issue, then handed back to the county’s own feed. Putnam was up and running eight minutes after posting the results our automated system missed. Westchester took longer. At first the orchestrator put the silence down to the scanner failures that had plagued the county for most of Election Day. Once it saw the site falling behind AP, it launched a subagent to investigate, which found the new results site. Westchester was up and running at 10:49 PM, about 45 minutes after the county first posted.

The rest of the night it was the other way around. We had Dutchess’s 10:14 update while AP still showed the old count. We had Rockland’s 12:10 update about 40 minutes before Decision Desk HQ. And when Westchester posted its first Election Day votes at 1:10 AM, we had them a minute later, before either desk.

the count vs. the desksNov 3 – 4
Ahead of the desks
Dutchess update
County posts 10:14›Our site 10:14›AP, still on the old count 10:18
Rockland update
County posts 12:10›Our site 12:11›DDHQ, still missing it 12:50
Westchester's first Election Day votes
County posts 1:10›Our site 1:11›AP last updated 12:50›DDHQ last updated 1:04
Fallbacks engaged
Putnam's results moved to a new address
Results up 9:38›Filled in from AP 9:44›County feed back 9:46
Westchester switched results sites
Results up 10:04›Filled in from AP 10:42›County feed back 10:49

What the numbers said

The story the site was built for turned out to be the story of the night. Heimish turnout didn’t dip the way it usually does in a midterm. By 1:10 AM the Heimish districts had cast 19,344 votes, more than their 2024 presidential-year total of 19,302. That’s 63% of their active voters, against 57% for the rest of Rockland. In 2024 it had run the other way: 63% against 70%.

That’s short of Lawler’s 30,000, though his figure was his own and took in Orthodox voters across the whole district, not only these 46 districts. The bet behind it held: the turnout came.

lawlervsconley.com — Heimish turnoutRockland
Heimish districts, votes counted19,344
2018
6,844
2022
15,592
2024
19,302

Already past 2024's 19,302 by 42, in a midterm.

Share of active voters who voted
Tonight
Heimish 63%
Rest 57%
2024
Heimish 63%
Rest 70%

They gave Lawler 97%, just as in 2024, and split almost evenly in the State Senate race. It just wasn’t enough to overcome the blue wave elsewhere. Outside the Heimish districts, Rockland swung about 12 points toward Conley. Dutchess swung about the same, Putnam about 15 and Westchester about 11.

At 10:36 PM our internal model’s own number was in: Conley finishes ahead by about 9,100 votes. That was its base case, which deliberately gave Westchester its 2024 lean, the cautious assumption. Minutes later the New York Times model put the final margin at Conley +9.4 points, so we checked the two against each other. Our base case came to about 3 points. Our swing case, with Westchester moving like Putnam and Dutchess had, came to about 8. The Times matched the swing case. The count, still not certified, has landed between our two cases: Conley by about 19,000 votes, or 6 points.

sanity check — final margin10:39 PM
Our model, base caseConley +3
Westchester votes like 2024
Our model, swing caseConley +8
Westchester swings like Putnam and Dutchess
New York Times modelConley +9.4
best estimate of the final margin
matchThe Times lines up with our swing case. By then, Putnam and Dutchess had both moved 8 to 9 points toward Conley, so the swing case looked like the realistic one.

By 10:15 we were confident Lawler’s path was effectively over, even though he had just taken the lead in the count. Outside the Heimish districts, Rockland was swinging hard toward Conley, Putnam and Dutchess were swinging further, and Westchester, his weakest county, hadn’t posted a vote. By 10:45, with Westchester’s first numbers in, we had the orchestrator drafting. Even if the Heimish districts reached their full 2024 turnout, he’d have needed about 58% of Westchester’s uncounted Election Day ballots, which was statistically highly unlikely in a county where he’d won 30% of the early and absentee vote. The swing and the model agreed.

We didn’t run a needle, and we’re not in the business of calling races. So at 10:51 PM we published strictly what the numbers supported: “Lawler’s path is narrow,” with exactly what he’d need and where it would have to come from. ABC News reported Lawler’s concession around 11:00. Our post went up at 11:05, and the site’s banner changed to “Lawler concedes” at 11:06, still saying, as it does, that no desk had called the race and votes were still being counted.

The night in numbers

~0
Page views, election night and the day after
0+
Unique visitors
~0
People on the site at once, at peak
0+
Visits from shared links
0+
Shares sent from the site
0
Posts after 8 PM
0
Written fully autonomously
0
Rockland districts tracked live

Everyone can build a tracker. What made this one worth building was everything a scraper can’t do: knowing which districts matter, checking that knowledge against the past, and having a team, human and AI, watching it all night with a clear line between what the agents decide and what people do.