Skip to content
Sangam Gupta
All work

DealShield AI

A Chrome extension that reads a tough negotiation email in Gmail and drafts three fact-checked replies you can insert in one click.

Status
Live
Stack
Chrome Extension (MV3), React, Vite, Tailwind CSS, Node.js, Express, Zod, Groq, gpt-oss-120b, Supabase, Razorpay, Render
Links
Add to ChromeWebsiteCode
reply strategies per email, each with the reasoning behind it
3
AI passes: draft, then fact-check against the email
2
free analyses a month; Pro is unlimited
10
  • DealShield panel in Gmail showing a prospect's email asking for a 33% discount, with an Analyze negotiation button
  • Three reply strategies, firm offer, trade-off and urgency, with the trade-off open showing what we give and what we get
  • The chosen draft inserted into Gmail's reply box above the signature
  • Fact-check result: an unsupported start date turned into a placeholder and an invented price rise removed

How it fits together

  1. Gmail thread
  2. Content script
    • Popup
    • Gmail sidebar
  4. Express API: auth and quota
  5. Groq gpt-oss-120b strategies
  6. Fact-check pass
  7. Insert into Gmail

Negotiation copilot for Gmail, published on the Chrome Web Store. A React extension reads the open thread, an Express API asks Groq's gpt-oss-120b for three strategies under a strict JSON schema, and a second pass checks every draft against the email. Free and Pro plans, with Razorpay subscriptions.

The problem

Replying to a prospect who wants 33% off, cites a competitor or sets a deadline is where deals are won or lost. A plain AI draft makes it worse in a new way: it invents deadlines, price changes or start dates you never agreed to, and you might send them.

How I built it

  1. Read the open Gmail thread with a content script; nothing leaves the browser until the user accepts a data notice and clicks Analyze, and the email text is never stored.
  2. Show the result in the popup and in a Shadow DOM sidebar inside Gmail, built with React 19, Vite and Tailwind on Manifest V3, asking only for storage and mail.google.com.
  3. Send the thread to an Express 5 API that checks the Supabase login, validates input with Zod, rate-limits, and enforces the Free and Pro quotas.
  4. Ask Groq's gpt-oss-120b for three strategies under a strict JSON schema generated from Zod: a firm offer, a trade-off that spells out the give and the get, and an urgency reply built from the prospect's own timeline. Retry once if a generation breaks the schema.
  5. Run a second pass that compares each draft with the email and turns anything it doesn't support, like a new deadline or price, into a [placeholder] for the user to confirm. If that pass fails, the drafts come back unchecked and the UI asks the user to review them.
  6. Insert the chosen draft into Gmail's reply box above the signature in one click.
  7. Keep error logs free of email content: only the status and error code are logged.
  8. Store subscriptions, usage and history in Supabase Postgres with row-level security; bill Pro through Razorpay subscriptions with signed webhooks.

Where it is now

Published on the Chrome Web Store with a website, privacy policy and support page. I built it with Claude Code as my pair programmer.

Ask about DealShield AI

Answers come from this case study and its code on GitHub.