ClearHelp

ClearHelp helps families under financial stress turn confusing food, housing, and school-support notices into plain-language checklists, clear next steps, and routes to nearby help.

Comment

Inspiration

ClearHelp was inspired by a parent or caregiver under financial stress who receives a school-meal, food-benefit, or housing notice filled with unfamiliar language and deadlines.

Help may exist, but finding and understanding it while stressed is difficult. A normal web search also assumes the user already knows what the notice means and what to search for.

What it does

ClearHelp turns a stressful notice into an actionable plan:

  1. The user pastes the notice.
  2. On-device AI identifies the likely support category.
  3. It extracts the deadline, requested action, and documents mentioned.
  4. It returns a plain-language explanation, checklist, confidence label, and immediate next step.
  5. The user can review a relevant support resource and request directions.

If confidence is low, ClearHelp does not guess or automatically recommend a resource. It returns a neutral “Needs review” result and asks the user to involve a trusted person.

ClearHelp also includes a secondary “get to help” feature. Its routing system can guide users to resources using walking routes that consider heat, shade, greenery, traffic noise, hills, and travel time.

Responsible AI by design

Risk: The classifier could misunderstand an unfamiliar notice or recommend outdated information.

Mitigation: Notice text stays on the device, confidence is described qualitatively, uncertain input receives a neutral result, and every synthetic resource is clearly labeled for verification.

Human in the loop: ClearHelp never determines eligibility, interprets legal rights, makes medical or safety decisions, or decides which documents must be submitted. A benefits worker, school employee, housing advocate, counselor, or trusted adult remains responsible for those decisions.

How we built it

The mobile application was built with Expo, React Native, and TypeScript.

The notice assistant uses a small on-device Naive Bayes text classifier trained with synthetic examples. Deterministic natural-language processing then extracts useful details such as deadlines, actions, and named documents. Because this processing is local, the notice does not need to be uploaded to an external AI service.

The routing backend uses Python, FastAPI, and NetworkX. It combines an OpenStreetMap street graph with:

  • SOLWEIG UTCI and direct-shadow GeoTIFFs
  • NOAA/DC LiDAR-derived elevation and tree-canopy data
  • DC land-cover data
  • DDOT traffic-volume data

A constrained routing algorithm compares feasible paths while respecting a maximum detour budget. Users can choose Balanced, Coolest, Green, Quiet, or Fastest routing styles.

The current support directory uses clearly marked synthetic demonstration records, as permitted by the challenge. These must be replaced with verified official listings before deployment.

Challenges we ran into

One challenge was making the AI helpful without allowing it to sound more certain than it really is. Early versions still showed a category-specific explanation for ambiguous notices. We changed the design so low-confidence input receives only a neutral human-review plan.

The geospatial backend introduced several challenges:

  • Processing very large one-meter environmental rasters
  • Keeping UTCI and shadow grids correctly aligned
  • Preventing incorrect place autocomplete results
  • Generating meaningfully different route profiles
  • Making live navigation responsive on a physical phone
  • Connecting an Expo Go device to a local backend through a VPN

We also found that limiting the number of returned routes could accidentally change the primary route. We separated search breadth from response count so the best route remains consistent.

Accomplishments that we're proud of

  • Built a complete input that allows AI processing which made helpful output
  • Kept notice analysis private and on-device
  • Added explicit uncertainty and human-review safeguards
  • Extracted notice-specific deadlines, actions, and documents
  • Connected the support plan to real walking directions
  • Used real UTCI and direct-shadow raster data in route scoring
  • Verified API route metrics against independent GeoTIFF sampling
  • Tested 32 valid API configurations and 24 invalid or boundary cases
  • Built live GPS navigation with rerouting, route progress, and compass rotation
  • Produced working Web, iOS, and Android builds

What we learned

We learned that responsible AI is not only a disclaimer. It changes product behavior. When the model is uncertain, the safest output is sometimes to stop classifying and bring in a human.

We also learned that finding support involves two different barriers: understanding what to do and physically reaching the right place. ClearHelp connects those steps while keeping the crisis-to-action translation as the primary experience.

Finally, we learned that environmental routing requires more than drawing alternative lines. The routes must be tested against their actual distance, temperature, shade, and greenery data.

What's next for ClearHelp

Next, we would:

  • Replace synthetic resources with verified 211 and official agency listings
  • Display source links, verification dates, and official phone numbers
  • Add camera scanning and OCR for paper notices
  • Support multiple languages and read-aloud accessibility
  • Evaluate the classifier with community organizations and diverse notice examples
  • Add a dedicated bicycle-network graph
  • Pilot the experience with parents, school family liaisons, and benefits navigators
  • Expand beyond Washington, DC while preserving local verification

Built With

Share this project:

Updates