Work / Case 01 — Georgia
The voters the file couldn’t see.
Georgia’s 2021 Senate runoffs turned on turnout at the margins. Standard voterfiles badly undercounted the state’s Asian American and Pacific Islander electorate — so we re-based the model, found the people who were there all along, and helped affinity groups reach them in a language and idiom that fit. The map was wrong. We fixed the map.
01 — The problem
A voter you can’t find is a voter you can’t reach.
Commercial voterfiles infer race and ethnicity from surname dictionaries and coarse Census geography. For the AAPI electorate — dozens of distinct communities, many with surnames the dictionaries miss — that method fails quietly. It doesn’t error out; it just returns a number that is far too small, and everyone downstream trusts it.
In a runoff decided by tens of thousands of votes, an undercounted community is a strategic blind spot. Field programs can’t turf what they can’t see. Affinity groups can’t call people who aren’t flagged. The voters were registered, active, and reachable — but the model rendered them invisible, so no program was built to talk to them.
02 — What we did
We re-based the model on how people actually cluster.
Instead of trusting surnames alone, we layered spatial clustering and homophily — the tendency of similar households to live near one another — onto the geo-referenced file. That surfaced high-confidence AAPI voters the dictionary approach never scored.
Start from geography, not guesses
Every household placed precisely, then read against Census and precinct reference.
Cluster on behaviour and affinity
Homophily analysis groups likely-similar households the surname model splits apart.
Score confidence, not labels
Hand programs a ranked, targetable universe instead of a flat yes/no flag.
Found isn’t the same as reached.
A bigger universe only matters if someone talks to it. We handed the re-based file to AAPI Victory Fund and the affinity groups already trusted inside these communities, and they ran the contact — by canvass, phone, and text, in the languages and idioms that actually land.
During early vote alone, that program reached roughly 65,000 voters who had never before been contacted in a culturally fluent way — people the old model had never even flagged. This is what polling as ethnography looks like carried all the way through to the door.
“Dheeraj makes the impossible possible. I come to him with a dream and he makes it into a plan.”
Varun — AAPI Victory Fund
Related
Start here