Maitraiyee Gautam

Stack: PyTorch, NLP, Longformer, BERT, AWS SageMaker, LangChain, LangGraph, CometML, Grafana.

• Trained HuggingFace Hydranet model via AWS SageMaker on 80K+ HTML e-receipts for structured extraction.
• Implemented LLM-driven validation using LangGraph and LangChain, automating rerouting of invalid receipts to Labelbox with error-tagged comments, reducing QA triage time by 30%.
• Deployed SageMaker inference endpoint managing over 100K receipts daily, validated with Locust benchmarking and shadow deployment across four environments.
• Boosted discount field accuracy by 21%, improving entity linking match rate by 2% in basket- and item-level audits.
• Monitored inference through Grafana, CloudWatch Logs, CometML, Slack Alerts, and PagerDuty to track volume, drift, and latency.
• Mentored junior engineers on best practices in machine learning, coding standards and agile engineering principles.

Stack: Python, deep learning, NLP, LangChain, RAG, vector search, SageMaker, Bitbucket Pipelines, Docker, Kubernetes.

• Trained the YOLOv8 model on over 60,000 receipt images using OCR and NER outputs to identify product group regions for entity linking.
• Automated data preprocessing and training workflows in SageMaker, enabling real-time monitoring with Comet ML, and CloudWatch.
• Deployed SageMaker inference endpoint, processing millions of receipts daily for over 12M users, including shadow traffic validation and A/B testing.
• Achieved 4% improvement in accuracy match and 2.1% gain in product group accuracy, validated against 4K+ audited receipts.
• Designed RAG pipeline utilizing LangChain and FAISS to retrieve answers from 500+ Confluence documents through semantic similarity.
• Deployed Mistral 7B on SageMaker GPU endpoint, delivering real-time completions with sub-second latency.
• Collaborated with DevOps to automate deployment of RAG microservice using Bitbucket CI/CD, Docker, and Kubernetes.
• Integrated Slack SDK, reducing documentation-related engineering queries by approximately 40%.

Stack: Python, AWS, Vector search, Kafka, Flink, Airflow, Kubernetes, CI/CD (Bitbucket Pipelines), Apache Hudi, Spark, EMR, Snowflake

• Engineered automated FAISS index update DAG, ingesting approximately 500 new retailers weekly to sustain current canonical store names.
• Enforced precision of 98% and recall of 95% using dynamic and static evaluations to block low-performing updates.
• Developed Kafka and Flink event-driven streaming pipelines, delivering real-time campaign metrics for advertiser optimization strategies.
• Reduced decision time from seven days to hours by providing low-latency data feeds deployed on Kubernetes.
• Constructed user purchase aggregate pipeline using SNS, Lambda, and SQS (DLQ) to process over 5 million daily purchase events with 99% delivery reliability.

Stack: Apache Hudi, Apache Spark (EMR), Amazon S3, Snowflake, AWS Glue, SQL.

• Developed batch ingestion backup for offer contributions utilizing Apache Hudi on S3 with Spark on EMR, achieving 60% reduction in outage recovery time.
• Integrated Hudi datasets into Snowflake, resulting in 35% decrease in ingestion costs and 40% enhancement in BI query performance.

Stack: Python, Pandas, NumPy, Scikit-learn, XGBoost, Jupyter Notebook, Matplotlib, Seaborn, Recommendation, and Ranking.

• Developed hybrid recommendation system to suggest optimal food categories for 200+ food banks in Feeding America network.
• Applied collaborative filtering to analyze usage similarity and identify relevant candidates.
• Trained XGBoost ranking model to score category-resource pairs based on cost, impact, fit.
• Evaluated model performance using Recall@K and NDCG, achieving 23% improvement in top-5 recall over baseline.

Stack: SSIS, SSRS, SSAS, MS SQL Server, T-SQL, Tableau, OLAP, Windows Server.

• Built data pipelines and reporting architecture using SSIS, SSRS, SSAS, and SQL to process and analyze data from more than 9 banks.
• Led reporting automation and OLAP cube design, reducing manual effort by 55%, and boosting regulatory reporting throughput 3x.