Justin Tang

Worked for NLP initiatives with transformers, RNNs, and sequence-to-sequence models for real-world language understanding and decision-support workflows

Architected highly available SaaS and production clinical AI on AWS. Grew from leading core product development to delivering RAG/NLP solutions over EHR data with strong SLOs, security, and measurable outcomes. Products: EHR CoPilot (RAG over EHR with citations), Risk Stratification Service (SageMaker XGBoost), Clinical SaaS platform (React + AWS), Glue/S3 data platform (millions/day ingestion). Key achievements Reduced tail latency by 80%: P99 improved from 1.0s to 200ms via cache-aside (Redis/ElastiCache), DynamoDB key/GSI redesign, and query refactors; typical P95 < 300ms. Delivered production RAG and clinical NLP: HIPAA/PIPEDA-compliant EHR CoPilot (Bio_ClinicalBERT + LLM, citations, safety filters) enabling ~40% faster clinician chart navigation. Shipped interpretable risk model: SageMaker XGBoost with Bayesian HPO achieving AUC 0.872 and precision@10% 0.713; pilot cohort saw a 12.7% reduction in ER visits. Implemented zero-downtime releases: Serverless backend on Lambda + API Gateway with blue-green and canary deployments, automated rollbacks, and 99.9% availability. Strengthened security and compliance: Cognito/Keycloak SSO, RBAC, KMS encryption, PHI de-identification, and audit trails; ~15% faster compliance reviews. Built data platform at scale: AWS Glue/S3/Athena pipelines ingesting millions of records per day; data contracts and lineage; BI queries P95 < 8s; internal freshness SLO T+2 hours. Established observability: Datadog APM/logs/traces with SLO burn-rate monitors; ~60% fewer noisy pages and 35–45% faster MTTR. Mentored team of three engineers: 18% reduction in critical bugs and 10% faster feature delivery through design reviews, testing strategy, and coding standards. Tech: AWS (Lambda, API Gateway, DynamoDB, MSK/Kafka, OpenSearch, S3, Glue, Athena, Step Functions), Node/TypeScript, Java, Python (ML), React/TypeScript, SageMaker, Vertex AI (TPU), Redis/ElastiCache, Datadog, OpenTelemetry, Terraform/SAM/CDK, Cognito/Keycloak.

Led a national-scale service desk for an energy major and an enterprise big-data platform for a top life insurer. Products: PetroChina Enterprise Service Desk (Spring/Oracle/Redis + CTI, 400k+ users, 99.95% uptime); New China Life Insurance Enterprise Data Platform (100-node Hadoop DW, 30M+ customers, 50+ TB). Built a 100-node Hadoop DW (HDFS/YARN) with Hive/Impala and Spark ETL via Oozie; standardized Oracle ingestion with Sqoop; served 50+ analysts. Cut reporting from 168h to 4h (~97.6%) and established a single source of truth. Improved ETL throughput ~60% via partitioning/bucketing, predicate pushdown, join plan tuning, and small-file compaction; enabled the first predictive policy-lapse model. Established governance and auditability: Kerberos/Ranger auth, Apache Atlas lineage, and data contracts with reconciliation to the system of record. For external reports, adopted a governed T+2 days publish window (automated validation, reconciliation, audit sign-off); internal marts remained near-real-time. Raised BI performance: migrated to columnar Parquet/ORC, enforced partition pruning, and added precomputed aggregates; tuned Impala resource pools for fairness; BI scan P95 < 8s. Delivered the PetroChina service desk: Spring Boot microservices with Oracle + Redis and WebSocket updates; Huawei AICC CTI (screen-pop, click-to-dial, skill routing). Achieved 99.95% uptime; chatbot deflected ~30% of Tier-0 tickets and reduced MTTR. Tech: Java (J2EE, Spring Boot), Oracle, Redis, WebSocket, Apache Hadoop (HDFS/YARN), Apache Spark, Hive/Impala, Parquet/ORC, Sqoop, Oozie, Apache Atlas, Apache Ranger, Kerberos, Prometheus/Grafana.

Research: Robotic vision systems with computer vision algorithms and sensor fusion Applied statistical modeling for decision-making in autonomous systems

Developed real-time control systems using C++/Advantech