experience
Production machine learning experience across evaluation, document intelligence, and applied NLP.
Production ML roles at Adobe, S&P Global, and American Express — from evaluation design and document AI to deployment and monitoring.
Adobe
Machine Learning Engineer Intern
- Built an LLM-as-Judge evaluation harness with LiteLLM to benchmark Gemini 2.5 Pro, GPT-4.1, and Claude 4.5 on multimodal chart-quality tasks, raising offline F1 from 79% to 92% by redesigning the scoring rubric against a human-labeled reference set and tracking every run in Weights & Biases.
- Shipped a LangGraph agent over Model Context Protocol and Adobe PDF Spaces that cut evaluation feedback from 48 hours to 30 minutes, with LangSmith tracing and a human approval gate on every generated plan.
S&P Global
Machine Learning Engineer II
- Fine-tuned and served LayoutLMv3 for token classification over financial filings on AWS EC2, versioning datasets and model artifacts with DVC and MLflow and promoting releases through CI/CD so training runs stayed reproducible and rollbacks were safe.
- Deployed Llama 2 on SageMaker and AWS Lambda with chain-of-thought prompt templates for page-level extraction, removing about 20 minutes of manual data collection per analyst task.
- Designed and deployed a T5 question generation and answering API on AWS EC2 that resolved current-year figures from long unstructured filings inside analyst-facing latency budgets.
- Owned the team MLOps workflow with DVC, MLflow, GitHub Actions, and Docker, and replaced a paid third-party parser with an in-house Java named-group regex service that saved roughly 10 minutes of analyst time per filing.
American Express
Data Scientist I
- Built and containerized VIBE, a production call-quality scoring model combining XGBoost with BERT embeddings, used as the system of record for annual representative incentive decisions.
- Deployed a three-tier complaint risk classifier with Naive Bayes, TF-IDF, and isotonic probability calibration on AWS EC2 so routing thresholds could be set on calibrated risk.
- Served an LSTM intent model with Flask on AWS EC2 that mapped call transcripts to demand types and prioritized incoming contacts, with PySpark feature generation over the transcript corpus.