Real-time & Batch Risk Scoring System
Company: TeleSign
Role: Machine Learning Engineer (2021-present)
TeleSign needed a scalable, multi-region Machine Learning solution for real-time fraud detection. The system had to process high-throughput user signals and deliver accurate risk scores with low latency, supporting telecom and identity verification use cases. Additionally, the team required a batch scoring capability for retrospective analysis and customer-wide risk evaluation.
Led the design and implementation of a dual-mode ML risk scoring system that supports both real-time inference and scheduled batch scoring. Developed end-to-end ML pipelines using AWS SageMaker for model training, evaluation, and registry integration. Deployed scalable inference endpoints with SageMaker capable of returning predictions in under 40ms. Built batch scoring workflows using AWS Glue and PySpark to support high-volume scoring on historical data. Designed feature engineering workflows that combine real-time streaming (Kinesis) and aggregated features from transactional databases. Automated multi-region deployment using Terraform and CI/CD pipelines (Jenkins) and integrated monitoring and alerting using CloudWatch to ensure system performance and reliability. The overall end-to-end latency for real-time scoring, including feature enrichment and orchestration, is approximately 300 milliseconds.
Tech stack:
ML/AI: scikit-learn, XGBoost.
Infra/MLOps: AWS SageMaker, Lambda, Glue, CloudWatch, Terraform, Jenkins.
Data Engineering: PySpark, AWS Kinesis, DynamoDB, Athena.
Languages: Python, SQL.
Deployment: Multi-region CI/CD, Docker.
Impact:
Reduced customer model delivery time by 70% through full pipeline automation. Achieved low-latency fraud detection with SageMaker inference under 40ms and overall end-to-end response time around 300ms. Enabled scalable batch scoring for large datasets across clients and geographies. Strengthened the security and reliability of telecom identity verification products.
RAG System for Code Documentation Retrieval
RAG System for Code Documentation Retrieval
Company: TeleSign
Role: Machine Learning Engineer (2021-present)
TeleSign’s engineering team needed a faster and more intuitive way to navigate internal code documentation. Traditional search tools were inefficient for large and evolving codebases, leading to lost time during onboarding, debugging, and implementation.
Developed and deployed a Retrieval-Augmented Generation (RAG) system tailored to internal codebases and documentation. The system indexes proprietary code, docstrings, and architectural notes by chunking and embedding them using Titan Embeddings. These embeddings are stored in OpenSearch, enabling fast semantic search. At query time, the top-k relevant chunks are retrieved and passed to Claude Sonnet via AWS Bedrock, which generates context-aware, natural-language answers.
The solution is fully serverless, with back-end logic implemented using AWS Lambda, ensuring scalability, cost-efficiency, and ease of deployment.
Tech stack:
AWS Bedrock (Claude Sonnet), Titan Embeddings, OpenSearch, AWS Lambda, Python.
Impact:
Reduced average documentation retrieval time by 20%, significantly boosting developer productivity and making technical references more accessible across the engineering team.
Smart Email Encoding for Loan Fraud Detection
Company: FICO
Role: Lead Analytic Scientist (2018-2021)
In application loan fraud detection, email addresses often contain subtle patterns that distinguish legitimate applicants from fraudsters. Traditional feature extraction methods failed to capture these sequential signals, limiting model accuracy and detection capability.
Developed a 1D Convolutional Neural Network (ConvNet) to process email addresses as sequences of characters, allowing the model to learn informative substring patterns automatically. The model encoded emails into meaningful embeddings that were then incorporated into the overall fraud detection pipeline.
After training, the TensorFlow model was serialized using Protocol Buffers (protobuf) and integrated into FICO’s proprietary Java-based inference pipeline, enabling production-grade real-time scoring.
Tech stack:
Python, TensorFlow/Keras, Scikit-learn, Java, Protocol Buffers, NumPy, Pandas.
Impact:
Achieved a 3% improvement in fraud detection accuracy for application loan risk scoring. The Deep Learning approach provided enhanced generalization for edge cases and helped reduce false negatives with minimal latency overhead in production.