VERA Ai

Inspiration

Only 15–18% of income-eligible households actually receive LIHEAP benefits in a given year, and research consistently points to procedural complexity — it's whether someone can actually read and act on a dense eligibility chart under stress, with real consequences riding on getting it right.

What it does

Two different flows:

Uploading a chart to Pre-fill fields on Web Form
Uploading a chart to Pre-fill an uploaded PDF form

A user uploads a photo of a chart or table they don't understand — an income bracket table, a benefits calculation chart, anything from their paperwork.

VERA:

Sends the image to a Agentic AI model, which extracts and renders structured data(household size, income, state, utility provider)
Autofills those values into a LIHEAP application form
Requires the user to verify every field before anything is finalized
Shows live progress while the completed form is generated
Ends with a completed application and a quick feedback check
Answers plain-language questions throughout — pulling from a built-in FAQ for common questions
Voice Based output

How we built it

Frontend: Vite Backend: Python FastAPI based server and LangGraph for Agentic AI that invokes Claude APIs.

Challenges we ran into

It took a while to study and leverage Langgraph to help orchestrate our agentic framework.
Trust vs. automation: deciding how much to let AI autofill without a mandatory human verification step, given the real-world stakes of a benefits application we landed on always requiring explicit confirmation before submission

What we learned

Real collaboration and delegation skills as a team despite a complex idea and approach.

AI accuracy and user trust are two separate problems. Early on we assumed the hard part would be getting the vision model to read charts accurately. It wasn’t, Claude handled that well. The harder problem was deciding how much to let it act on what it found. A wrong autofill on a real benefits application isn’t a minor bug, it’s a household’s energy assistance at risk. That’s why we built the verification step as mandatory, not optional every extracted field has to be confirmed by a human before it goes anywhere. We learned that for high-stakes dom

What's next for VERA Ai

Build the real FastAPI backend for PDF generation and genuine job-status progress

Expand beyond LIHEAP to other means-tested programs with similarly dense paperwork — SNAP, Medicaid, FAFSA

Add multilingual support for LEP users, and audio/screen-reader-friendly chart narration for blind and low-vision users

Partner with community organizations or caseworkers who already work with these populations, to test accuracy and usability with real applicants rather than synthetic data

Run a small accuracy study on the vision-extraction step, since a false field value on a real benefits application is a meaningfully different risk than a wrong answer in a chatbot