Skip to content
Back to selected work
Exhibit B

Eureka

AI scholarship matching and mentorship platform.

Lead Full-Stack Engineer (Contract)May 2026 – Present

app.myeureka.app
Eureka matching scholarships to a studentScholarships are scattered in a vector space. A student profile appears, hard eligibility gates remove the ineligible ones, dense retrieval pulls the nearest toward the student, and LLM scoring settles them into three fit tiers while the rest fade.Tier 1Tier 2Tier 3Student1 Gates2 Dense retrieval3 LLM scoringScholarships in vector space

What I built

AI-powered scholarship-matching and mentorship platform for international students (myeureka.app).

  • Designed a multi-stage matching pipeline: hard eligibility gates (destination, gender, age), dense vector retrieval in Qdrant (chosen over a hybrid BM25/RRF design after an nDCG A/B, 0.918 vs 0.844) and LLM scoring into three fit tiers.
  • Built an agentic ETL pipeline that uses LLMs to extract structured eligibility criteria, deadlines and quotas from scholarship sources, with A/B evaluation of extraction quality.
  • Implemented authentication end to end: email verification, Google OAuth, TOTP, refresh-token rotation, password reset, parental consent for minors, and audit logging.
  • Built an asynchronous notification system on Celery and RabbitMQ (transactional email with circuit breaker, deadline reminders, web push) and mentor bookings with Google Meet links.
  • Integrated Stripe Checkout, subscriptions, dunning, session credits and mentor payouts, with webhook handling and money analytics.
  • Deployed to production on DigitalOcean with OpenTelemetry tracing, Prometheus metrics and Grafana Cloud, and resolved memory and latency issues in production.

Built with

  • Python
  • FastAPI
  • SQLAlchemy
  • PostgreSQL
  • Qdrant
  • Celery
  • RabbitMQ
  • Redis
  • React
  • TypeScript
  • TanStack
  • Stripe
  • Docker
  • OpenTelemetry
  • Grafana