Data & Problem Solving
Signals in messy real world systems turned into results.
Voter Signature Verification Pipeline
Finding 5,000 in a pool of 13.5 million
The challengeA political campaign had thousands of petition signatures challenged by the Board of Elections and needed to demonstrate, on a tight court-ordered deadline, that enough of those signatures belonged to real, eligible voters. I was brought in to build the system that checked each disputed signature against New York's official voter records and produced a clear, defensible spreadsheet the legal team could submit.
The hard part was not the volume of data, it was the messiness of it. People sign petitions by hand. The same person might write "Abbie" while the state has them on file as "Abigail." Someone might list a new address after moving, use a maiden name, drop a middle name, or simply make a typo. Meanwhile, the official New York voter database holds 13.5 million records, including thousands of people who genuinely share the same name.
So the real question for every signature was deceptively simple but technically tricky: is this handwritten name actually a registered, eligible voter in New York, and can we tell which specific person it is and not someone else with the same name?
Rather than relying on one rule, I matched each signature using a ladder of verification methods, from strongest evidence to weakest, and labeled every result with a confidence level:
- Exact identity match using the voter's name and state-issued voter ID together.
- Unique-name matching, where an uncommon name appears only once in the state and can stand on its own.
- Address-based tiebreaking for common names, so two different "Maria Rivera" records get separated by matching the petition address to the address on file.
- Fuzzy matching to catch nicknames, misspellings, and dropped middle names that an exact match would miss.
- Borough and location narrowing to resolve ambiguous cases down to the most likely individual.
Crucially, anything I could not confidently confirm was never silently dropped. It went into a clearly labeled review queue so the legal team retained full visibility and made the final calls themselves.
Every contested signature ran down the ladder, strongest evidence first. Nothing was silently dropped.
The result- Delivered the complete verification package on deadline, the core deliverable I was responsible for.
- Roughly 4,000 signatures had direct, high-confidence confirmation against the New York voter database, with another 1,379 subject to further manual review, each carrying a plain-English label explaining how the match was made and how strong it was.
- Produced a full audit trail documenting every match decision and its reasoning, so any individual result could be traced and explained.
- Processed a 6.2 GB voter file (13.5 million records) efficiently on a single machine, backed by an automated test suite to keep results consistent under time pressure.
Skills: Large-scale data processing · Entity resolution and record matching · Fuzzy / approximate string matching · Data cleaning and normalization · Building defensible, auditable outputs for non-technical stakeholders · Delivering under a hard deadline.
Spatial Equity at the GSD
Gund 1970; before space was a luxury
A financial model demonstrating a sustainable path for the GSD to create a better student experience.