Envis - Crisis To Action Translator

Home Page
Personal Info For Better AI Context
Clear And Easy To Read Document Summary
Reminders And Todos
Annotations On Documents

Inspiration

Everyone has felt it: a hospital bill, an eviction notice, a benefits letter, or discharge instructions lands in front of you — dense, intimidating, and written for the institution that sent it, not for you. In that moment people don't need a 40-page explainer. They need someone calm beside them who says "here's what this actually means, and here's exactly what to do next."

We built ENVIS to be that someone — a crisis-to-action translator that turns the scariest piece of paper in your life into clear, plain-language steps, without making you feel small for not understanding it.

What it does

You drop in a document — a photo, a PDF, or pasted text — and ENVIS:

Reads it like a detective and gives a calm, structured breakdown: the bottom line, what stands out (flagged 🔵 key / 🟡 watch-out / 🟢 good), a line-by-line verdict on charges, what's missing, and your exact next moves.
Annotates the original image — OCR + AI highlight the real lines on your document so you can hover and see what each one means.
Localizes everything — it infers the country, currency, and laws from the document and judges costs/rights against local norms (e.g. Indian hospital schemes like Ayushman Bharat vs. U.S. charity care).
Knows your rights — a research agent looks up the institution's policies and your protections.
Acts for you — it drafts the email/letter to send, builds reminders and to-dos for real deadlines, finds support near you and benefits you may qualify for, and exports a shareable one-pager (PDF + QR).
Meets you where you are — replies translate into 10 languages, with a larger-text accessibility mode and a "Night Calm" theme.

Crucially, ENVIS explains and suggests — it never makes a legal, medical, or financial decision for you. You stay in control.

How we built it

Frontend: Next.js (App Router) + TypeScript + Tailwind, with GSAP for the calm, breathing micro-interactions.
AI advisor: OpenAI GPT-5 family behind a heavily engineered system prompt that forces reasoning over the actual numbers instead of generic checklists, plus tool-calling to create reminders/to-dos.
Documents: Tesseract.js runs OCR in the browser; a second model maps each line to a plain-language meaning and an importance color to drive the highlights.
Data & auth: Supabase (Postgres + Auth) with row-level security so every user's threads, messages, reminders, and profile stay private.
Agentic research: a LangGraph agent powers the "My rights" briefing.
Extras: MapTiler for the 3D "support near you" map, and client-side PDF/QR generation for the one-pager.

We also score how much simpler we made the text, using a Flesch-style reading-ease score shown before → after:

$$\text{ease} = 206.835 - 1.015\left(\frac{\text{words}}{\text{sentences}}\right) - 84.6\left(\frac{\text{syllables}}{\text{words}}\right)$$

Challenges we ran into

"Is it stuck?" latency. GPT-5 at high reasoning effort gates its first token on finishing a long thinking phase, and we were also forcing tool-calls on the first turn — a second full model pass before any text appeared. We fixed it by streaming the analysis token-by-token, tuning reasoning effort for a fast first token, and moving reminder/to-do creation to a non-blocking pass after the reply renders. Time-to-first-word dropped from a minute-plus of dead air to a few seconds of live, streaming text behind a document-aware progress trace.
Aligning OCR with AI. Tesseract's noisy line boxes had to line up with the model's analysis so highlights landed on the right text — lots of merging, filtering, and importance-tagging.
Localization without assumptions. Getting the model to reason about local prices, laws, and aid programs — and to clearly signal confidence when estimating — instead of defaulting to U.S.-centric advice.
Big payloads & privacy. Base64 document images were too large for ordinary DB columns, and RLS meant being deliberate about exactly what each request could read or write.

What we learned

With reasoning models, perceived speed is a product decision: streaming + effort tuning + decoupling side-effects matters as much as the answer itself.
The hardest part of "AI for vulnerable moments" isn't the model — it's the tone, the responsible-use guardrails, and the accessibility that make someone feel seen, clear, capable, and in control.