Work

Municipal governments sit on decades of as-built drawings, handwritten service records, and construction documents that are trapped in paper or unstructured PDFs — invisible to the GIS systems staff rely on every day.

A suite of computer vision and AI pipelines that extract structured data from these documents at scale — from vision language models that read handwritten water service cards to automated as-built ingestion that pulls metadata, organizes files, and feeds results directly into ArcGIS Pro catalogs, with a human-approval gate before anything enters the catalog of record.

Decades of paper became a living, queryable system — from scanned drawing, to georeferenced GIS feature, to natural-language answer. Higher-quality infrastructure data with traceable lineage back to the source record. Building it stretched me across computer vision, model training, cloud architecture, agentic AI, and RAG.

Over a century of handwritten water service records were trapped on deteriorating paper cards — unsearchable and disconnected from modern systems — while field crews kept capturing new records in inconsistent formats.

One initiative, two directions. Backward: a local vision-AI model reads the handwritten cards and extracts structured, searchable data at scale. Forward: a structured Survey123 + Field Maps intake form so crews capture clean, GIS-ready records from here on.

A century of infrastructure knowledge made searchable and integrated into GIS — roughly 40+ hours of manual entry completed in about 4 hours of review — plus a standardized capture process that keeps the archive clean going forward.

A lakefront suburb needed residents and visitors to understand Central Business District parking at a glance — including the question everyone asks, “how far is this lot from where I'm going?” — and off-the-shelf ArcGIS Experience Builder couldn't deliver that experience.

I designed, built, and deployed a custom React + ArcGIS application: every CBD lot mapped with color-coded restrictions and detailed profiles, on-demand walk-time routing from any user-set origin, and a mobile-first interface — all driven by a single JSON config that makes it portable to other communities.

I owned it end to end — requirements, stakeholder data validation, build, and deployment — with no internal template to work from. Because I built it config-driven, a second municipality was served from the same codebase with no code changes.

Public crime data is often locked in formats no resident can use — and publishing it openly raises real privacy concerns about pinpointing incidents to a single household.

Built a Python pipeline that ingests incident data, anonymizes each record to the 100-block level to protect privacy, and auto-publishes a clean public dashboard to ArcGIS Online — replacing a manual update process with a repeatable, scheduled one.

What took 60–90 minutes of manual work per update now runs in about 5 — giving residents a transparent, privacy-respecting, always-current view of local incidents while reclaiming staff capacity.