Skip to content
Tony Nguyen

Experience

Production graph and streaming infrastructure for fraud prevention at Uber

I interned at Uber through UberSTAR in 2023, came back for a second internship in 2024, and joined full-time after graduating. All three roles were on fraud-detection infrastructure, turning high-volume event streams into the graph signals and model features that risk systems act on. Dollar figures are the impact of my projects.

Company
Uber
Span
–
Roles
2 internships, then full-time
Domain
Fraud and risk
Infra
  • Kafka
  • Cassandra
  • Pinot
  • Flink
  • Hive
  • Spark

01Roles

  1. 01Fraud detection · Identity and risk graphs

    Software Engineer · Uber

    Dates
    to
    Kind and duration
    Full-time · 1 yr 6 mo
    Location
    San Francisco, CA
    Teams
    Risk Knowledge Platform (uFlow + uGraph)
    Knowledge Graph Platform (uGraph)

    Built and extended the graph infrastructure behind Uber’s fraud-prevention systems, connecting identities, organizations and documents into signals that risk models and enforcement can act on. I owned work from data modeling through streaming pipelines to production rollout.

    • Identity graph

      Extended Uber’s identity graph to unify user identities and risk signals across third-party platforms, supporting partnerships representing $600M in incremental gross bookings.

      • $15M+ in fraud-prevention impact
    • Enterprise-risk integration

      Designed and deployed an enterprise-risk integration that expanded verification and challenge-history modeling from individual customer organizations to complete enterprise hierarchies, enabling graph-derived fraud features.

      • $10M+ in measured fraud losses prevented
    • Document-risk graph

      Led the document-risk graph migration from single-market to multi-jurisdiction status modeling for drivers and couriers worldwide, increasing actionable fraud-signal coverage by 30%.

      • $5M+ in measured fraud reduction
    • Event semantics

      Productionized change-aware event semantics across all streaming graph pipelines, suppressing no-op updates and distinguishing newly created from updated state, which improved the reliability and efficiency of downstream Kafka processing.

  2. 02Bot detection · Real-time graph analytics

    Software Engineer Intern · Uber

    Dates
    to
    Kind and duration
    Internship · 4 mo
    Location
    Sunnyvale, CA
    Team
    Knowledge Graph Platform (uGraph)

    Built real-time graph analytics for detecting automated traffic, from the streaming ingestion path to the query layer that serves bot signatures under a tight latency budget.

    • Bot-detection graph analytics

      Built real-time bot-detection graph analytics processing 8K peak events/sec into up to 40K graph writes/sec across millions of security-signal relationships.

      • $4M+ in measured fraud-prevention impact
    • Graph ingestion and query

      Built graph ingestion and query layers that retrieve the top-100 bot signatures from endpoint vertices with 1M+ connected edges at 200 ms end-to-end latency.

      • $1M in annual downstream compute-cost savings
    • Temporal graph model

      Designed Kafka and Cassandra streaming data flows and a temporal graph model across URL endpoints, JA3 TLS fingerprints and request headers, enabling rolling access-count analytics over 1- to 24-hour windows.

  3. 03ML infrastructure · Point-in-time features

    UberSTAR Software Engineer Intern · Uber

    Dates
    to
    Kind and duration
    Internship · 4 mo
    Location
    Sunnyvale, CA
    Team
    Risk Applied Machine Learning

    Built point-in-time feature infrastructure for fraud models, so training data reflects what was actually known at decision time instead of what is known today.

    • Point-in-time features

      Built reusable point-in-time ML infrastructure that expanded an initial fraud model’s candidate feature set by 58% and improved recall by 2%.

      • $5M+ in modeled annual fraud prevention
      • ~$200K in measured annual impact
    • Historical feature access

      Productionized historical feature access for fraud-model pipelines at 92K writes/sec, eliminating a 3+ month feature-collection delay and accelerating offline training and experimentation.

02Earlier experience

  1. to Cupertino, CAPart-time

    Computer Science Teaching Assistant, De Anza College

    • CIS 22A — Beginner Programming Methodologies in C++
    • CIS 40 — Introduction to Programming in Python
  2. to Internship

    Software Engineer Intern, CodeDay

    • Contributed to Winston, a logging library used in 1M+ repositories.
  3. to San Jose, CAPart-time

    Coding Instructor, Code Ninjas

    • Mentored children ages 5–14 in programming, including JavaScript, Python and Scratch.

03Education

University of California, Berkeley

B.A. Computer Science ·