Technical troubleshooting
Move from symptoms to a clear problem statement, then give the customer a practical path forward.
AI evaluation · Software security · Geospatial analysis
Technical judgment,
made useful.
I build tools for AI evaluation, investigate software security, and analyze geospatial and environmental data. My work spans source provenance and correction-aware testing, raster analysis, and rainfall patterns. I turn complex technical questions into reproducible workflows with evidence that can be inspected and challenged.
Discovered stored cross-site scripting in Appointment Hour Booking (CVE-2026-96573).
01 Build reproducible AI evaluations
02 Investigate software security
03 Analyze geospatial and climate data
Professional experience
Technical support with clarity and follow-through.
At Starlink, I troubleshoot technical issues, use AI-assisted workflows to make support clearer and more consistent, and help peers adopt effective practices through training and process documentation.
AI-assisted workflows
Use AI to structure complex cases, strengthen response clarity, and turn effective working methods into repeatable support practices.
Peer training & documentation
Help colleagues adopt effective workflows through practical coaching and clear process documentation.
Career path
From service and field leadership to technical systems.
-
Sep 2025—present
Starlink customer support · SpaceX
Technical troubleshooting, AI-assisted customer-support workflows, peer training, and process documentation.
-
2022—2024
Field management & coaching · FPI
Frontline leadership, coaching, and consistent execution across field work.
-
2018—2022
Service work · The Salvation Army
People-centered service in environments where clarity and follow-through matter.
Independent research
Research with the evidence attached.
Public preprints and an interactive example make the reasoning available for inspection.
Interactive demonstration
Correction after a handoff.
A deterministic example shows how two records can share a compressed decision but require different responses when the same premises change.
How the example works
Two fictional rules preserve different arrangements of supporting reasons. Withdrawing the same premises changes their decisions differently. This constructed example illustrates information loss in a handoff; it does not measure language-model performance.
Selected research preprints
Selected public contributions
Evidence before adjectives.
Published security research, open-source patches, and inspectable analysis tools—each linked to its source.
Published security advisory
Wordfence advisory · 30 September 2026
Appointment Hour Booking stored DOM-based XSS
Discovered stored cross-site scripting in Appointment Hour Booking (CVE-2026-96573). The finding affects versions through 1.5.97 under a non-default configuration.
Included in v3.2.0
KeeperHub · v3.2.0 · Pull request #2340
Generation-stream failure handling
I contributed to generation-stream error handling and cleanup included in KeeperHub v3.2.0.
Built on Yoh5’s original diagnosis and accepted plan.
Contribution notes
The patch separated JSON parsing from deliberate stream errors, canceled still-open readers on early failure, preserved the original error if cleanup failed, and added mocked-stream tests.
Patch contribution
LibTIFF · Merge request !946
Directory-tracking memory accounting
I contributed a patch routing directory-tracking allocations through TIFF handle-aware allocators, together with a dedicated unit test.
Maintainer Su Laus adapted the approach in MR !992 and credited Zack.
Contribution notes
The test_open_options tests in the maintainer’s
MR !992 came from the separate test contribution in MR !975.
Public prototype
EvidenceDesk · Strands + local Ollama/Qwen
A research handoff that keeps the evidence attached
I designed and directed a local research workflow that captures source snapshots, checks exact textual matches and exports attributed review packets.
Text matching supports quotation checks; interpretation remains a reviewer’s task.
Public evaluation package
Vesuvius · TIFXYZ resampling
Checking coordinate-resampling validity
I evaluated and corrected coordinate-resampling validity on three public surface patches from one scroll, with regression tests and source-data checksums.
The report covers direct API geometry checks, separate from ink recognition and upstream integration.
Interactive work sample
NOAA GHCN-Daily · Rainfall comparisons
Rainfall, in context
An offline rainfall explorer that makes accumulation windows, reference choices and missing-data assumptions inspectable. Charts and numerical tables retain unavailable comparisons instead of silently changing the selected reference.
Methods in brief
The explorer compares dated accumulation windows against selected reference members. It keeps a reference unavailable when required dates fall outside the frozen extract. A separate synthetic masking exercise tests the effect of replacing withheld rain with zero.
xarray · Scientific Python documentation
Submitted a concrete Dataset.reindex_like example showing coordinate-label alignment and how missing labels become NaN.
GIS & environmental analysis
A geographic way of seeing systems.
B.S. in Geography and Environmental Studies, California State University, Northridge — May 2025, Cum Laude. My upper-division coursework includes advanced environmental GIS, medical geography, and remote sensing.
Working toolkit
- ArcGIS Pro & Online
- Remote sensing · Landsat
- Raster analysis
- Spatial overlays
- Zonal statistics
- Python for PRISM acquisition
- Excel
Māori Land in Aotearoa New Zealand
Changing Landscapes of Customary and Freehold Land
I built an ArcGIS StoryMap combining regional land-ownership mapping, a historical swipe comparison, and parcel-scale exploration to explain changes in Māori land in Aotearoa New Zealand.
StoryMap contents
The published story includes three web maps, a regional percentage view, an 1860/2017 North Island swipe comparison, and exploration of remaining customary-land parcels including Te Neke. It identifies Māori Land Dataset 2017 and New Zealand Region Boundaries 2021.
The maps show historical land-ownership data, with source datasets credited in the StoryMap.
View the StoryMapMapping Hurricane Harvey’s Urban Rainfall Extremes
I automated PRISM precipitation downloads with Python, clipped August 2017 monthly rasters to Texas and Harris County, and used county-level zonal summaries to explore spatial variation.
Cartographic methods
I applied cubic resampling and Jenks classification for display. August 2017 monthly totals include rainfall outside Hurricane Harvey; resampling changes presentation, not measurement resolution.
Rethinking Wildfire Risk
Hydroclimate whiplash, urban growth, and dynamic fire hazard mapping in Los Angeles County
I created thematic maps in ArcGIS Pro comparing the Eaton and Palisades fire perimeters with shrub/scrub land cover and CAL FIRE hazard zones, supported by a Los Angeles rainfall chart.
Methods and source notes
I clipped NLCD shrub/scrub land cover (class 52) to Los Angeles County and used Raster Calculator to create a binary raster. I overlaid CAL FIRE and LA County GIS layers, assembled map layouts, and visualized LA Almanac rainfall records in Excel.
These are descriptive overlays; terrain, wind exposure and fuel moisture were not modeled.
Climate Change & Flood Risks: The ARkStorm Case Study
Co-presented with Shaïli Azerad, this literature-based case study explained California atmospheric-river flood risk and the ARkStorm planning scenario.
Contact