Reply Wise

ReplyWise Approve-to-Reply AI Email Assistant

Comment

ReplyWise — Project Story

Inspiration

ReplyWise was inspired by a very real and personal problem:
replying to emails is often stressful, risky, and harder than it looks.

We realized that the hardest part of email communication isn’t typing — it’s decision-making:

  • What does the sender really want?
  • Is this email urgent or risky?
  • Am I accidentally agreeing to something?
  • How do I reply confidently in a language that is not my native one?

Most AI tools today focus on speed and automation.
We wanted to focus on safety, accessibility, and control.

That’s why we designed ReplyWise around a simple principle:

AI should assist human judgment, not replace it.

What We Built

ReplyWise is an AI-powered browser extension that lives directly inside your inbox.
There is no copy-paste, no context switching, and no auto-sending.

At a high level, ReplyWise:

  • Analyzes incoming emails
  • Detects potential risks (urgency, financial requests, commitments)
  • Generates multiple reply styles (short, friendly, formal)
  • Supports bidirectional translation (English ↔️ Turkish)
  • Enables voice interaction for listening and controlling replies
  • Requires explicit user approval before sending

Mathematically, you can think of ReplyWise as an AI-assisted decision function:

[ \text{Final Reply} = \text{Human Approval} \circ \text{AI Suggestion} ]

If approval does not happen, the action is canceled:

[ \text{Send} = \begin{cases} 1, & \text{if user approves} \ 0, & \text{otherwise} \end{cases} ]

This guarantees a human-in-the-loop design.

How We Built It

Architecture Overview

ReplyWise uses a lightweight, modular architecture designed for scalability and safety.

Frontend

  • Angular 21
  • Angular Material
  • TypeScript, SCSS
  • Runs as a browser extension embedded in the inbox UI

Backend

  • Node.js + Express
  • TypeScript
  • Zod for runtime validation
  • Helmet & CORS for security

AI & Voice

  • Google Gemini 2.5 Flash
    • Email intent understanding
    • Risk detection
    • Reply generation
  • ElevenLabs API
    • Natural text-to-speech for emails and drafts
  • Web Speech API
    • Voice command recognition

Voice Interaction Model

ReplyWise supports voice-based control using a simple command state machine:

  • “Approve” / “Send it” → proceed to send
  • “Edit” / “Change it” → open editing flow
  • “Reject” / “Don’t send” → cancel the process

This allows hands-free and accessibility-first interaction without sacrificing safety.

Accessibility & Language Support

One of the core design goals of ReplyWise was inclusivity.

We implemented:

  • Text-to-speech for incoming emails and generated replies
  • Voice commands for navigation and approval
  • Translation support:
    • English email → Turkish reply
    • Turkish email → English reply

This makes ReplyWise especially helpful for:

  • Non-native English speakers
  • International students
  • Neurodivergent users
  • Visually impaired users
  • High-stress or hands-free environments

Accessibility was not an add-on — it was designed from day one.

Challenges We Faced

1. Balancing Automation vs. Control

One of our biggest challenges was resisting the temptation to fully automate replies.
Auto-sending is easy — safe decision-making is not.

We solved this by enforcing explicit approval at every critical step.

2. Voice Reliability

Voice recognition can be inconsistent depending on environment and browser support.
We designed fallback paths so that every voice action has a click-based alternative.

3. Risk Detection Without Overreach

We needed to flag risky emails without overwhelming users with false alarms.
This required careful prompt design and structured validation.

4. Time Constraints

Building a polished extension, backend, AI logic, and voice features within a hackathon timeline was challenging.
We prioritized core experience and correctness over unnecessary features.

What We Learned

  • Responsible AI is a design decision, not a feature
  • Accessibility should be considered at the architecture level
  • Lightweight AI systems can be powerful when combined thoughtfully
  • Judges and users value clarity and trust more than flashy automation

Most importantly, we learned that good AI products slow users down at the right moments.

What’s Next

If we continue building ReplyWise, next steps include:

  • Outlook and Microsoft Teams integration
  • Enterprise policy rules for compliance
  • Team-level risk analytics
  • Support for more languages
  • Deeper personalization while preserving privacy

Final Thoughts

ReplyWise is not about replying faster.

It’s about helping people:

  • Think clearly
  • Communicate safely
  • Feel confident
  • Stay in control

AI assists. Humans decide.

Share this project:

Updates