Duy Tien Nguyen
Backend Engineer
I own the auth service on a six-service platform being rebuilt off a monolith. Every internal request authorizes against it. Permission checks went from 32ms to under 2ms, and invalidation runs on events rather than TTLs, so a revoked permission can't outlive its revocation.
Skills
Experience
SMART GENERATION OF DIGITAL
Ho Chi Minh City, Vietnam- Own
auth-service, the IAM layer every other service authorizes against, through a monolith-to-microservices rebuild: data model, caching, failure behaviour. - Cut permission checks from 32ms to under 2ms with a three-tier read cascade: CASL matrices precomputed in Redis, falling back to Neo4j then MongoDB, still correct when a tier is cold. (K6, 350 req/s, staging)
- Invalidated cached permission matrices on events published through a Transactional Outbox rather than on TTL, so a revoked permission never survives in cache; consumers keyed on event UUID make retries idempotent.
- Projected the MongoDB organisation graph into Neo4j as a read model synced by change streams, keeping recursive role inheritance traversable, with cycle detection blocking circular reporting lines during org restructuring.
- Encrypted customer data at field and file level and contributed the key-wrapping layer of the platform's Rust PKCS#11 key service, where master keys never leave the HSM; also implemented mTLS and JWKS rotation between services.
Selected Projects
Second Prize, University Research Competition • Prometheus, Grafana, Kafka, MinIO, Nginx, Docker
- Instrumented a full backend stack behind a single dashboard, exposing system load and Kafka consumer-group lag through custom exporters, clearing processing bottlenecks logs never showed.
Advanced Algorithms (individual project) • Python, Monte Carlo simulation, hypothesis testing
- Benchmarked five bandit policies against round-robin and least-connections across workloads (150 Monte Carlo trajectories, T=10,000), cutting mean request latency 51.9ms to 18.6ms.
- Reported two results that contradicted my own hypotheses: policy rankings reverse under different tuning budgets, and forgetting mechanisms never paid off at any drift density tested.
Reinforcement Learning (course project) • Python, Discrete CQL, PyTorch, Numba, Dynamic Programming
- Built a hybrid Safe-RL system: a Conservative Q-Learning policy wrapped in a hard safety layer, capturing savings at 0% constraint breaches.
Anchored IPFS content hashes of donation evidence on-chain, making the audit trail tamper-evident rather than trust-based.
Education
HO CHI MINH CITY UNIVERSITY OF TRANSPORT
Expected Dec 2026Bachelor of Science in Data Science — GPA 3.86/4.0
Honors: Eureka semi-finalist (2025) • Merit Scholarship, Top 1%
Certifications: Google Cybersecurity Professional (2025) • IBM DevOps & Software Engineering (2026)