Groundtruth: FEMA Damage Assessment Platform
Phone photos in, FEMA-ready damage records out, with AI that separates every photo and a person who confirms every item
At a glance
I acted as design lead and executive sponsor for Groundtruth. I studied real FEMA records from Hurricane Michael, wrote the platform design and the government AI guardrails, designed every screen, and took it from prototype to a working test build on AWS, using Claude (AI) to help build it.
The challenge
After a disaster, local governments must document every damaged facility to FEMA's standards. Today the same facts get typed into five forms and hundreds of photos get sorted by hand, and state AI rules require that any AI help stays labeled and approved by a person.
What I led
- Studied real FEMA records: a DR-4399 Site Inspection Report, a Project Worksheet, the Damage Inventory and PDA templates and the Photo Sheet
- Mapped the workflow from field capture to the DDD and Scope of Work that FEMA reviewers expect
- Researched the market: TerraFort, Fulcrum, ArcGIS Survey123, Veoci and Inspectial
- Wrote the AI guardrails from the Louisiana AI Acceptable Use Policy and Executive Order JML 25-103
- Designed the platform, the photo review flow, the new-file wizard and trial runs, for phone, tablet and desktop
- Built and deployed a working test build on AWS with photo AI on Amazon Bedrock
- Set up a demo workspace and an in-app feedback loop for SME validation
The outcome
A working test build runs on AWS: photos go in from a phone, AI separates them into items, a person confirms each one, and all five FEMA documents fill from the same records. Next is SME validation with real damage photos, then a pilot with one applicant.
Overview
Groundtruth is a FEMA Public Assistance damage-assessment platform. Inspectors capture photos on a phone, AI separates each photo into items, an office reviewer confirms every one, and the platform fills the PDA, Damage Inventory, Photo Sheet, Site Inspection Report and the Damage Description & Dimensions and Scope of Work from the same records. Repair costs compare a city’s own history in QuickBooks with RSMeans and Xactimate.
How I led it: I acted as design lead and executive sponsor. I studied real FEMA records from Hurricane Michael (DR-4399) in Panama City, wrote the platform design and the government AI guardrails, designed every screen, and took it from prototype to a working test build on AWS. I used Claude (AI) to help build the prototype and the app.
What’s in the test build:
Photo separation: AI lists each item in a photo with a box, count, condition and visible damage, and a person confirms, edits or removes it
Five FEMA documents from one record: PDA, Damage Inventory, Photo Sheet, Site Inspection Report and the DDD and Scope of Work
Three cost sources: QuickBooks history, RSMeans and Xactimate side by side, with the reviewer’s choice logged
New-file wizard and trial runs: seven steps to a new file, plus a 10-step checklist for practice runs
Government AI guardrails: a data notice, an on/off switch, sensitive-photo flags and an AI use log
Phone, tablet and desktop: camera capture on site, review in the office
Built to government AI requirements: every AI feature follows the Louisiana AI Acceptable Use Policy and Executive Order JML 25-103, with labeled AI output, a named human approver, no AI eligibility or pricing decisions, a data notice and off switch, protection of personal information, data checks and an AI use log
Services
Customer & Market Research, Product Strategy, Government AI Governance, UX & Interaction Design, Interactive Prototyping, AI-Assisted Development, Cloud Deployment (AWS), Go-to-Market Planning
Client name
Groundtruth (independent product by John Karl)
Team
John Karl, Design Lead & Executive Sponsor
Prototype
Market research
User research
UX Decision
Go-to-market
This is a working prototype. Follow the guided steps or switch tabs and use the controls. Everything responds live.
Market research
Goal:
Find out what agencies use for FEMA Public Assistance damage documentation today, what it costs and where it leaves them stuck.
TerraFort: advertises offline voice capture, remote sensing, AI-generated DDDs, scopes and estimates, a GIS map builder and weather intelligence. A seven-day trial and custom enterprise pricing are advertised; accuracy data and human-review controls are not published.
Fulcrum: a Public Assistance damage report template for Categories A–G that works offline and can be shared with FEMA for desktop review. Enterprise pricing on request; G2 reviewers mention rising prices and limited reporting.
ArcGIS Survey123 and Field Maps: the basis of FEMA’s PDA survey templates, with strong maps and offline forms, included in ArcGIS licensing. G2 reviewers cite a steep learning curve for non-GIS users and basic built-in reporting.
Veoci: FEMA PA workflows with damage forms, force-account cost logging and PDF exports formatted to FEMA templates. Pricing by demo.
Inspectial: advertises AI damage assessment for inspections.
The gap: no tool ties every photo, measurement and price back to the FEMA forms with AI that a person must confirm item by item, inside state AI rules.
User research
Goal:
Learn how FEMA Public Assistance damage is documented today from the real records, find where time and accuracy are lost, and pin down the rules the product must follow before anyone uses it on a live disaster.
What the FEMA records showed (real input)
Site Inspection Report, Hurricane Michael (DR-4399), Panama City: 13 handwritten pages for one Category D damage record covering 8 drainage-ditch segments. Every segment’s length, width, depth and begin and end GPS were written by hand, then typed again into the Damage Inventory and the damage description. One segment was measured from the roadway because the inspector could not enter the property, yet the form has no place to say a number is estimated. Most mitigation, insurance and environmental answers were marked Unsure, one EHP answer (work within 200 feet of a waterway) was Yes, and the applicant’s representative initialed every page.
FEMA Project Worksheet (Category E contents): a tabletop review with no site visit, built from the applicant’s documents. The damage description lists 88 items one line at a time in a fixed pattern, “Contents, 6 each of Wooden Pews, 8.66 LF long, cause, 0% work completed,” and the scope of work repeats the same 88 lines as “Replace 6 each of…”. Costs came from internet pricing and applicant quotes for a $107,699.79 project at a 75% federal share, with an insurance review that set an obtain-and-maintain requirement.
Damage Inventory template: one row per damaged facility with category, location, latitude and longitude, specific damage, primary cause, approximate cost, percent complete, labor type, prior assistance and priority, with fixed pick lists for Categories A–G, labor types and Yes/No/Unsure.
Preliminary Damage Assessment template (FEMA Form 90-81): a site estimate per location with category, GPS, damage description, impact and cost, rolled up into a summary by category, plus written statements on public health and safety and economic impact.
Photo Sheet: four photos per page, each needing a damage description and photo number under a header with the Damage Inventory number, category, work order, FIPS code, GPS and inspector.
Rules Groundtruth must follow (real input)
Louisiana AI Acceptable Use Policy: AI output must be labeled, reviewed and approved by an identified person before use; AI may not make program-eligibility or financial decisions on its own; users must opt out of model training; AI data must follow records retention.
Executive Order JML 25-103: agencies must keep an inventory of AI use cases, review and clean data before it goes into an AI system, and keep personal information and specific property information out of AI tools until an approved policy covers them.
Requirements carried in (real input)
TerraFort competitive requirements review: six personas, rules that must never be bypassed (length of damage is not the replacement quantity; an estimated measurement never becomes observed; generated text is never an eligibility decision; corrections never overwrite the field record) and pilot gates: at least 30% less documentation time, 95% of critical fields correct, zero unsupported AI statements in approved output and 100% human sign-off.
Proto-personas (hypotheses)
Field inspector: captures sites and photos quickly, often with weak signal, and needs to know a shot is good before leaving
PA project specialist: turns evidence into the DDD and scope of work and needs every number traceable
QA reviewer: resolves gaps and double counts before anything goes to FEMA
Applicant finance lead: needs cost sources it can defend in an audit
State AI and records officer: needs to see what the AI did, who approved it and where the data lives
Research plan: SME review of the demo workspace with in-app feedback on every screen; timed trial runs on the same DR-4399 records against today’s paperwork to test the 30% time target; AI accuracy measured from the AI use log (accepted, corrected, rejected, missed); then a pilot with one applicant on a live disaster.
Why Each UX Decision Was Made
Goal:
Meet the pilot gates from the requirements review: zero unsupported AI statements and 100% human sign-off in anything sent to FEMA, with a target of at least 30% less documentation time than the paper process in the DR-4399 records.
Capture a fact once, fill five forms: the Panama City SIR shows each segment’s dimensions and GPS written by hand and then retyped into the Damage Inventory and damage description, so one site record now fills the PDA, Damage Inventory, Photo Sheet, SIR and DDD/SOW.
Every number says observed or estimated: one SIR segment was measured from the roadway with no way to flag it, so every measurement and quantity carries a basis tag, and estimated values print with an ESTIMATED or QTY TO VERIFY mark (the requirements review’s rule that estimates never become observed).
AI items use the Project Worksheet’s own pattern: the worksheet lists 88 items as count, item, size, damage and percent complete, so photo separation returns exactly those fields and confirmed items drop straight into the DDD line format and a matching “Replace” scope line.
A person confirms every item: the state policy requires an identified human owner for AI output, so every AI item starts as Needs check, uncertain items need an extra “I checked this in the photo” tick, and the reviewer’s name is stored with each decision.
Photos carry their own evidence: the worksheet was a tabletop review with no site visit, so each photo keeps GPS, time, shot type and who took it, and the Photo Sheet fills its description and photo number from that record.
Begin and end shots lock the location: the SIR records begin and end GPS for every segment, so the shot list asks for Begin and End and the site fills its coordinates from the first geotagged photo.
Answer the SIR questions in the app: most mitigation, insurance and environmental answers in the SIR were Unsure, so each answer is recorded with the applicant’s initials, an Unsure insurance answer raises a review flag, and a Yes on the 200-foot waterway question adds an EHP note to the scope.
Pick lists match the FEMA templates: the Damage Inventory template fixes categories, labor types and Yes/No/Unsure answers, so the app uses the same lists and the export matches the template column for column.
AI never prices anything: the policy bars AI from financial decisions and the worksheet’s costs came from internet pricing and quotes, so a price comes only from QuickBooks history, RSMeans, Xactimate or a quote, and the source is logged per line.
Notes are context only: in testing, the AI listed vegetation in a photo just because the inspector’s notes mentioned it, so the instructions now say to list only what is visible, which serves the zero-unsupported-statements gate.
Notice, off switch and use log: the executive order calls for an AI use-case inventory and limits what goes into AI tools, so each file has a data notice someone must accept, an off switch, sensitive-photo flags and an AI use log.
Go-to-Market Plan
Goal:
Earn subject-matter-expert trust first. The test build is live, so the first 90 days recruit PA experts to validate the demo, then one applicant to pilot it on a real disaster.
Positioning: for local governments and recovery consultants who must document disaster damage for FEMA, Groundtruth turns phone photos into FEMA-ready records with AI a person confirms item by item
First audience: FEMA PA consultants and city and county public works and emergency management teams in hurricane states
Offer: a guided 30-minute demo, then a private trial-run workspace with the 10-step checklist. Use the Request a trial button at the top of this page
Pilot gate: SME sign-off on the demo, user sign-in in place of the test access code, and the agency’s own AI and IT review before any real records go in
Pricing: not set. A per-applicant or per-disaster license will be tested during the pilot
90-Day Rollout
Days 1–30, validate: 5 to 8 PA experts work through the demo and trial runs and leave feedback on each screen
Days 31–60, harden: fix what SMEs flag, add user sign-in and file exports, and measure AI accuracy from the AI use log
Days 61–90, pilot: one applicant runs a real file from capture to package alongside their current process
AI that follows government rules
Before the first photo goes to the AI, someone reads a short data notice and turns Assist on, and their name is recorded. Agencies without an approved AI policy can leave it off and work by hand. Assist flags photos that show faces, license plates or personal documents and never identifies anyone. It checks photos for blur and missing scale references, and flags an item seen in two photos as a possible double count.
Every AI action and every human decision on it lands in the AI use log, with model and region. That log feeds the agency’s AI use-case inventory and shows how often the AI was accepted, corrected, rejected or missed something, which is the accuracy measure SMEs use to judge it.
Government AI compliance: the system is designed to follow the state’s AI requirements and protocols end to end: AI content is labeled, a named person owns every AI-assisted decision, AI never decides eligibility, funding or price, models do not train on agency data, data quality is checked before it reaches the model, and every AI use is inventoried. Each agency still runs its own IT and AI review before go-live.
Groundtruth test build: one field record fills the PDA, Damage Inventory, Photo Sheet, Site Inspection Report and the DDD and Scope of Work, with every AI-drafted line labeled until a named person approves it.
Final observations
Groundtruth shows how I lead a product into a regulated space. I started from real FEMA records and the state’s AI rules, turned the rules into design decisions, and shipped a working test build that subject-matter experts can use with their own photos. AI does the sorting and drafting; people stay accountable for every item and every approval. Next steps are SME validation, user sign-in and file exports, then a pilot with one applicant on a real disaster.
John Karl
19 Years • Product & UX Design Leader Leading Mobile Product for 500K+ Users
If you like what you see or have any questions, feel free to send me an email anytime.
Selected work
[2010 -2025]
