Stellanet

Stellanet helps students find matched professors and draft better outreach emails using Amazon Nova on AWS, with transparent ranking, tone rewrite, and human-in-the-loop review.

Comment

Inspiration

Stellanet was inspired by a common challenge for students and early researchers: finding the right faculty mentor and writing meaningful first-contact emails is time-consuming, unstructured, and often discouraging. I wanted to build a practical AI system that makes outreach more focused, transparent, and effective without removing human control.

What it does

Stellanet helps users run an end-to-end research outreach workflow:

  • Takes user interest, profile, and target universities.
  • Retrieves and ranks relevant faculty candidates using grounded research signals.
  • Shows why a match is relevant with explainable details.
  • Generates personalized outreach drafts.
  • Rewrites drafts by tone (professional, friendly, short).
  • Keeps users in control with human-in-the-loop review before sending.
  • Supports secure auth flows including email/password and Google sign-in.

How I built it

I built Stellanet as a full-stack web app:

  • Frontend: React + Vite + Tailwind for a responsive workflow UI.
  • Backend: FastAPI with modular services for auth, discovery, and draft rewrite.
  • AI layer: Amazon Nova 2 Lite (via Amazon Bedrock) for ranking/drafting/rewriting.
  • Grounding source: OpenAlex API for research and affiliation signals.
  • Auth/Session: local SQLite in development, production-ready PostgreSQL support.
  • Deployment: Dockerized backend on AWS App Runner and frontend on Vercel.

I also implemented environment-based configuration, API abstractions, and route handling improvements for production reliability.

Challenges I ran into

  • Production deployment differences between local/dev and cloud environments.
  • Health check failures caused by invalid database host configuration.
  • OAuth origin mismatch issues in Google Cloud setup.
  • Route refresh behavior on production domains for SPA paths like /signin and /about.
  • Double-slash API URL edge cases causing endpoint 404s in production.
  • Ensuring fallback behavior remained user-safe while still being debuggable.

Accomplishments that I am proud of

  • Built a complete, usable outreach product instead of a model-only demo.
  • Integrated Amazon Nova into the core user value path.
  • Delivered grounded, explainable ranking and controllable rewrite workflows.
  • Implemented secure auth with Google sign-in support.
  • Deployed backend on AWS App Runner with containerized delivery.
  • Improved UX reliability with route-safe refresh handling and robust error pathways.

What I learned

  • AI product quality depends as much on reliability engineering as model quality.
  • Small configuration mismatches (env vars, hostnames, trailing slashes) can break core flows.
  • Grounded generation + transparent reasoning improves trust and usefulness.
  • Shipping with iterative debugging and real user-style testing is critical for production readiness.
  • Building with Nova is strongest when tightly integrated into a practical workflow, not just a chat interface.

What's next for Stellanet

  • Improve ranking calibration and confidence transparency.
  • Add persistent workspaces, saved runs, and draft history/versioning.
  • Expand personalization controls for outreach style and intent.
  • Add stronger observability and automated quality checks.
  • Introduce advisor/team review workflows for collaborative outreach.
  • Explore multimodal inputs (CVs, statements, and research docs) for richer context.

Built With

Share this project:

Updates