Epione

Inspiration

I was nine the first time I had to read an eviction notice out loud. My mom handed it to me at the kitchen table because I was the one who read English. I didn't know what "eviction" meant, I just knew, from her face, that I was supposed to. Millions of kids do this; sociologists call it language brokering. We built Epione so no child has to be the family's lawyer, and no parent has to wait for one.

What it does

Maria, a Spanish-speaking mom, gets a confusing official letter. She texts a photo to a normal phone number and seconds later gets back, in her language: what it is, the exact deadline with a live countdown, and 3 steps plus a real help line. Then it does what a chatbot never would, because Epione is an agent, not a chat wrapper: it connects her to local programs (searches help near her ZIP, checks what she qualifies for, and with her okay starts the intake and hands her to a real caseworker), drafts her reply, flags scams before she pays, and reminds her before the deadline. It works on any phone, no app, no data plan, because the people who need this most are least likely to own the newest phone.

How we built it

Next.js 16 on Vercel. AI SDK v6 with a provider-flexible model layer (auto-detects Claude / OpenAI / Vercel AI Gateway). Vision + Zod-validated structured extraction (document type, deadline, amount, action, verbatim source quotes, scam score, read-confidence). Confidence-gated tiered routing: reads on a cheap model, escalates to a frontier model only on low-confidence reads. Deadline math in code (deterministic countdown). Two channels: Twilio (SMS/MMS) and a free Telegram bot, plus a browser demo. Encoding-aware (GSM-7 vs UCS-2) SMS cost optimization. Remotion for the pitch video.

Challenges we ran into

Making it an agent, not a chat wrapper, the hard part was getting the AI to act (match local programs, check eligibility, connect a human). SMS economics: carrier A2P 10DLC takes weeks, so we made the demo channel-agnostic (web + free Telegram) while keeping SMS as the vision; we also learned SMS encoding drives cost more than the model. And doing it responsibly: a wrong deadline could cost someone their home.

Accomplishments that we're proud of

It works end-to-end: on a real eviction notice it pulled the landlord, the $1,450 owed, the deadline, computed the days left, and quoted the source lines at 95% confidence, and it flags a gift-card scam as high-risk. ~5¢ per family helped at scale. A genuinely agentic flow, not just Q&A.

What we learned

The people who most need help with official mail are least likely to have a smartphone or English, so the channel (SMS) is the whole equity story. Language brokering is a real, widespread, lived problem. And an agent that finishes the job beats a chatbot that answers a question.

What's next for Epione

Real 211 / legal-aid / benefits-screening API integrations, A2P 10DLC registration for the live SMS line, more languages, and a pilot with a city or nonprofit.