Responsibilities include:

• I am part of the Microsoft Defender of Office (MDO) Security Research Org at MSFT, specifically working on building out the machine learning capabilities and scalable infrastructure of the product. I am also part of the Sonar Machine Learning (Sonar ML) team, which is the full detonation platform we built for detonating threat vectors in real time.
• Built an internal small language model (SLM) for Business Compromised Emails (BEC), Spam, and Phish detection, optimizing perception DNNs in FP16/INT8 precision for reduced computational overhead and enhanced production efficiency using NVIDIA’s TensorRT and CUDA.
• Developed and deployed a near real-time computer vision model for detecting and decoding malicious QR codes in messages, saving over 25 million dollars annually in COGs.
• Created an ONNX Model Predictor library in C# and .NET for CPU inference, allowing sub-30 millisecond real-time inference across multiple Microsoft organizations as a NuGet package.
• Lead engineer for ensuring infrastructure compliance with MSFT Security First Initiative (SFI), including authentication, cloud storage security, and network isolation for VMs.

Projects include:

• Implemented RLHF (Reinforcement Learning from Human Feedback) pipelines in Langchain, LlamaIndex, and Langraph for personalized therapy solutions at Cartha, including custom retrievers and query transformers.
• Applied model optimization techniques (pruning, quantization, kernel fusion) for low-latency edge applications and implemented RAG systems with NVIDIA acceleration for enhanced inference performance.
• Optimized RAG systems using hybrid vector+sparse retrieval, implementing ColBERT, HyDE, and CoT for enhanced reasoning of multi-step queries and developed custom tokenizers and embedding models for specialized security datasets.
• Fine-tuned and deployed open-source language models (Phi, Llama, Mistral), leveraging the Unlsoth library for LoRA and quantization methods for edge deployment.

Responsibilities included:

• Designed scalable AI infrastructure using Terraform and CloudFormation, leveraging MLOps principles with CI/CD via GitHub Actions.
• Built high-throughput data pipelines using AWS Glue, Apache Spark, and S3, and automated ETL workflows with AWS Step Functions.
• Deployed scalable ML workloads using AWS Lambda and EKS, and built real-time data ingestion systems with Apache Kafka and Amazon Kinesis.
• Developed containerized ML services with Docker, managed deployments using Helm charts, and configured Prometheus and Grafana for system health monitoring.