MyJellyBean

MyJellyBean turns suspicious messages into a risk score, red flags, and a safer reply—so you act with clarity, not panic

Comment

Inspiration

Scammers and impersonators often succeed by creating urgency and confusion, which makes it hard to think clearly in the moment.
So we built MyJellyBean: bite-sized clarity for high-pressure messages—a fast “second set of eyes” that highlights red flags and suggests calmer, non‑escalatory next steps.

What it does

MyJellyBean is a full-stack message safety analyzer for suspicious DMs, texts, and emails. You paste a message (optionally add platform/context + risk toggles) and it returns:

  • A risk score (0–100) and risk category (scam/fraud, impersonation, harassment/abuse, coercion/manipulation, privacy risk, meetup escalation risk, etc.)
  • The top signals/red flags that drove the score
  • A “Do this now” checklist with non-escalatory actions (verify, preserve evidence, report, block, don’t share OTP)
  • A safer reply draft to avoid oversharing
  • A structured report summary that’s easy to copy into platform reporting workflows

How we built it

  • Frontend: React + Vite + Tailwind for a fast, clean UI
  • Backend: Express (Node.js) API endpoint that accepts the message + context
  • AI layer: Google Gemini via @google/genai, prompted to return strict JSON
  • Safety guardrails: We explicitly avoid retaliation/doxxing guidance and emphasize de-escalation, verification, and reporting
  • Deployment: Hosted on Vercel with GEMINI_API_KEY stored as an environment variable

Challenges we ran into

  • Reliability: Getting consistent structured outputs required strict JSON prompting and robust fallback behavior when outputs don’t parse.
  • Safety UX: Writing “helpful but non-escalatory” advice is tricky; we tuned the wording to avoid confrontation while still being actionable.
  • Deployment details: Ensuring secrets stayed server-side and were correctly configured as environment variables in production.

What we learned

  • Building around an LLM is as much about constraints and validation as it is about model quality.
  • Small UX choices (clear risk bands, concise checklists, copy-to-clipboard) make the tool feel more usable in stressful moments.
  • “Human safety” features need explicit guardrails so the product helps users respond safely, not just detect risk.

Built With

  • TypeScript
  • React
  • Vite
  • Tailwind CSS
  • Node.js
  • Express
  • Google Gemini API (@google/genai)
  • Vercel

Built With

Share this project:

Updates