the problem
Every food crisis follows the same pattern: applications spike, food banks report shortages, a news story runs, everyone scrambles. By then, families have already gone weeks without adequate food.
No government has a system that says: "financial stress in these zip codes just spiked 40%. Here are the programs that should be surging outreach. Here's who is affected and what to do." They find out the same way everyone else does — after the damage is done.
what freshfind does three steps
FreshFind gives governments a head start on food crises. It works in three steps:
Example: Financial stress spikes 35% in eastern Kentucky. The system flags that 2,400 households are at risk, mostly adults 55-64, nearest fresh food 12 miles away. The state SNAP director gets a briefing with recommended actions. The agency sends targeted outreach. Families get help before the crisis peaks.
how the AI actually works the data is real
The Inform step is the core of the grant — and the hardest part. Here's how it works, and why we're confident it can.
The AI doesn't generate predictions from thin air. It cross-references fast-moving signals against deep contextual data to produce specific, actionable briefings. Not every source updates daily — and that's by design. The real-time layers (Google Trends, FEMA, Federal Register) are the triggers. The slower layers (Census, BLS) provide the context that makes those triggers actionable. Every source below is either already integrated into the Barometer or publicly available through federal APIs.
Update cadence: Google Trends: daily · FRED: daily to monthly · BLS: monthly · Census: annual. The Barometer blends these into a composite score that updates as fast as its fastest input.
How we use it: The Barometer already aggregates this into zip-code-level stress scores across all 50 states. The AI layer adds anomaly detection — when a score spikes beyond normal seasonal variation, it triggers the pipeline.
Update cadence: Federal Register: daily · FEMA declarations: real-time · State bulletins: as published. These are the fastest-moving trigger signals in the system.
How we use it: NLP pipeline classifies every change by program, affected population, geography, and effective date. A new SNAP work requirement in Kentucky doesn't just get flagged — it gets matched to the specific households it will affect.
Update cadence: Census ACS: annual · SNAP QC: annual · FNS enrollment: monthly. This data changes slowly — demographics don't shift overnight. Annual updates are appropriate for the context layer role these sources play.
How we use it: When a stress spike or policy change is detected, the system overlays demographic data to estimate: how many households, what age groups, what program eligibility. This turns a signal into a population estimate.
Update cadence: USDA Atlas: updated periodically · SNAP retailers: quarterly · Google Maps: real-time · Feeding America: quarterly. Store-level access is the most dynamic piece — Google Maps keeps this current.
How we use it: Every briefing includes a food access score: can people in the affected area actually reach a store that carries fresh food and accepts EBT? This is the difference between "people need SNAP" and "people need SNAP and the nearest grocery is 12 miles away."
Update cadence: Ongoing research relationship. Calibration data refreshed with each new pilot community. This is the correction layer — it doesn't need to be real-time, it needs to be right.
How we use it: This is what calibrates the model. Public data says Wolfe County, KY has 4 grocery stores. Our research says 2 of them don't carry fresh produce. That calibration is the difference between a useful system and a naive one.
The output isn't a dashboard or a score. It's a structured briefing for a specific government decision-maker:
Generated: March 17, 2026 · Confidence: High
Signal: Financial stress index +35% over 30-day baseline (zip codes 41301, 41311, 41339)
Cause: Expanded SNAP work requirements effective April 1 (Federal Register 2026-04291)
Affected: Est. 2,400 households · 60% adults ages 55-64 · 73% current SNAP recipients
Food access: Nearest fresh produce 12.3 mi avg · 2 of 4 stores carry produce · No public transit
Recommended actions:
1. Activate targeted outreach to affected SNAP recipients in 41301, 41311, 41339
2. Coordinate mobile pantry deployment (Feeding America contact: on file)
3. Alert 211 center for anticipated call volume increase
4. Deploy FreshFind eligibility screener for alternative program enrollment
Sources: BLS Local Area Unemployment · Census ACS 2024 · USDA SNAP QC · Federal Register · FinMango Barometer · UVA field research
Every claim in the briefing traces back to a specific data source. The AI assembles and contextualizes — it doesn't hallucinate. Recommendations follow pre-approved templates that government partners help define.
This isn't speculative. The data sources are public and machine-readable. The Barometer already aggregates financial stress data from 5 federal APIs. Gemini's context window handles the cross-referencing. The UVA research provides the calibration layer that most government AI projects lack.
What the grant builds is the pipeline that connects these sources and the briefing engine that turns raw signals into decisions. The AI doesn't need to be novel — it needs to be reliable, auditable, and fast. That's an engineering problem, not a research problem.
See the full technical architecture — model pipeline, infrastructure, and compliance →
try it yourself live tools
The response tools are already built and working. The grant connects them to the intelligence system.
Financial Health Barometer
Real-time financial stress tracking across all 50 states. This is the data that powers the Detect step.
Food Desert Analyzer
Check whether a neighborhood has real access to grocery stores and fresh food.
Food Assistance Calculator
One intake screens for SNAP, WIC, school meals, food pantries, and more.
built on real research
Two years of field research with the University of Virginia across five communities. We mapped who's food insecure, why programs aren't reaching them, and what's blocking access. These communities are our pilot partners.
Chicago, IL
~500k food insecure. $300M SNAP distributed monthly. Dense infrastructure but severe neighborhood-level access gaps.
Atlanta, GA
Only 38% of SNAP retailers in majority-Black neighborhoods carry produce vs. 75% elsewhere.
Kansas City, MO
~79k food insecure adults. Cross-border wage disparity creates compounding economic stress.
Wolfe County, KY
27.3% food insecurity. 4 food stores in the entire county. This is where the model was stress-tested.
Read the full technical overview, government use cases, roadmap, and measurement framework →
Governments shouldn't have to wait for a crisis to find out it's happening. FreshFind sees the problem forming, tells the right people, and puts food on the table before anyone has to ask.
Scott Glasgow · Sarah Cherian · Soham Patel · Tony Ramos
FinMango · finmango.org
