Outsync

Outsync is a full-stack AI web application that generates personalized academic outreach emails using a stateful LangGraph pipeline and Gemini 2.5 Flash. The system takes a user’s resume and research/professor URLs to produce tailored drafts based on tone and length preferences.

---

Overview

Academic outreach emails are hard to get right:

• they need to be short but specific,
• personalized without sounding generic,
• and aligned with your actual background.

Outsync solves this by combining:

• agentic multi-step generation (LangGraph),
• Retrieval-Augmented Generation (RAG) over your resume,
• and structured refinement to improve the final draft.

---

Key Features

• Resume-grounded personalization (RAG)
  • Upload a resume and retrieve relevant experience/projects during generation.
• URL-based context extraction
  • Provide professor/research URLs and extract key information to ground the email.
• Stateful multi-step workflow
  • LangGraph pipeline supports structured drafting + refinement.
• Tone and length controls
  • Configure the output style (e.g., formal vs casual, short vs detailed).
• Draft improvement loop
  • Generates an initial draft, then refines for clarity, relevance, and alignment.
• Secure authentication
  • Google OAuth with secure sessions.
• Persistent user data
  • MongoDB stores user preferences and configuration state.
• Export-ready drafts
  • Copy/export the final email from the UI.

---

Tech Stack

LLM + Orchestration

• LangGraph (stateful agent workflow)
• Gemini 2.5 Flash (generation + refinement)

Retrieval / RAG

• ChromaDB (vector store)
• Embeddings
  • sentence-transformers/all-MiniLM-L6-v2 (Hugging Face)
  • (Resume version notes: Cohere embeddings were also tested/used depending on configuration) :contentReference[oaicite:0]{index=0}

Backend

• FastAPI
• MongoDB
• Google OAuth
• Deployed on Render (backend hosting)

Frontend

• Next.js + Tailwind (generated via V0.dev)
• Deployed on Vercel

Data Extraction

• Firecrawl (scraping professor/research pages)

---

System Design

Outsync is designed as a modular LLM application with three core layers:

1. Frontend UI

   • Handles resume upload and configuration inputs (tone/length)
   • Displays generated drafts and supports export

2. Backend API

   • Auth + session management
   • Stores user configuration state in MongoDB
   • Executes LangGraph workflow and returns results

3. LangGraph Workflow

   • Multi-step pipeline that produces higher-quality outputs than a single prompt call

---

LangGraph Workflow

Outsync uses a multi-node LangGraph pipeline with a state object passed through each stage.

Example workflow stages:

1. Extract Professor/Research Context
   • Scrape URLs and summarize relevant details (lab focus, publications, keywords)
2. Resume Retrieval (RAG)
   • Embed resume content and retrieve top relevant chunks for personalization
3. Draft Generation
   • Create an initial outreach email draft using retrieved resume context + professor info
4. Refinement
   • Improve clarity, conciseness, tone alignment, and specificity
5. Final Output
   • Return export-ready email draft

---

Retrieval-Augmented Generation (RAG)

To prevent generic emails and improve personalization, Outsync uses RAG over the user’s resume:

• Resume is parsed into chunks
• Chunks are embedded using:
  • sentence-transformers/all-MiniLM-L6-v2 (Hugging Face)
• Embedded chunks are stored in ChromaDB
• At generation time, the system retrieves the most relevant resume sections using cosine similarity
• Retrieved context is injected into the prompt to ground the output

This improves:

• factual accuracy (no made-up experiences),
• relevance to the professor’s work,
• and overall specificity.

---

Authentication and User State

Outsync supports user accounts via Google OAuth.

Persisted data includes:

• user profile info (basic identity)
• saved prompt configurations
• generation preferences (tone, length)
• session state and prior drafts (optional depending on deployment config)

MongoDB is used to persist user data and configuration state.

---

Deployment

Outsync is deployed as two services:

Backend (FastAPI)

• Hosted on Render
• Responsible for:
  • OAuth/session handling
  • RAG + LangGraph execution
  • MongoDB persistence

Frontend (Next.js + Tailwind)

• Hosted on Vercel
• Responsible for:
  • UI + upload workflows
  • configuration controls
  • displaying and exporting drafts

---

Local Development

Prerequisites

• Python 3.10+
• Node.js 18+
• MongoDB (local or cloud)
• API keys:
  • Gemini API key
  • Google OAuth credentials

---

Backend Setup (FastAPI)

# create environment
python -m venv venv
source venv/bin/activate

# install dependencies
pip install -r requirements.txt

# run backend
uvicorn app.main:app --reload

Frontend Setup (Next.js)

npm install
npm run dev

Environment Variables

Example backend .env:

GEMINI_API_KEY=your_key_here

GOOGLE_CLIENT_ID=your_google_client_id GOOGLE_CLIENT_SECRET=your_google_client_secret OAUTH_REDIRECT_URI=http://localhost:8000/auth/callback

MONGODB_URI=your_mongodb_connection_string

CHROMA_PERSIST_DIR=./chroma_store

Example frontend .env.local:

NEXT_PUBLIC_API_BASE_URL=http://localhost:8000 NEXT_PUBLIC_GOOGLE_CLIENT_ID=your_google_client_id

Example Usage

Sign in with Google

Upload your resume

Paste professor/research URLs

Select tone and length preferences

Generate draft

Refine/export final email

What I Learned

Building Outsync reinforced a few key lessons:

Multi-step workflows outperform single-prompt generation for quality and consistency

RAG is essential for grounding personalization in real user data

Production LLM apps require good state handling, not just model calls

A strong UI for configuration and iteration is critical for usability

Future Improvements

Add multi-professor batching (generate outreach emails for multiple targets at once)

Add evaluation harness (quality scoring, hallucination checks)

Support more document types (CV, cover letter, transcript)

Add prompt versioning + draft history

Add stronger citation grounding from extracted professor pages