Priyanka Deshpande

About Priyanka Deshpande

AI/ML & Data Science professional with 10+ years of experience building and deploying scalable, production-grade machine learning systems that drive real business impact. I specialize in delivering end-to-end solutions from data engineering and feature development to model deployment and MLOps across complex, real-world environments. My expertise spans classical ML, NLP, and Generative AI, with hands-on experience in designing robust models, RAG pipelines, and LLM-powered applications. I focus on building reliable, interpretable, and high-performing systems by emphasizing data quality, evaluation rigor, and overfitting prevention. I have strong experience in Azure-based architectures, developing low-latency APIs, real-time and batch pipelines, and implementing MLOps practices like experiment tracking, CI/CD, and model monitoring. I’ve also worked extensively on reducing hallucinations in LLMs through retrieval strategies, guardrails, and structured evaluation frameworks. I enjoy solving complex problems, collaborating cross-functionally, and translating data into actionable insights that improve efficiency, reduce risk, and enable smarter decision-making. Passionate about combining strong ML fundamentals with modern Generative AI to build scalable, responsible, and impactful AI systems.

Experience

1. Senior Machine Learning Ai Engineer

   TD

   Oct 2023 — Present · US

   Project: Scalable Computer Vision & GenAI Model Service Tech Stack: Python, TensorFlow, FastAPI, Docker, MLflow, AWS, OpenAI API (LangChain), PostgreSQLProject Overview: Developed an end-to-end production-grade system for fine-grained image classification and integrated a Generative AI layer to provide automated, context-aware nutritional insights and reporting based on visual inputs.Roles and Responsibilities:• Architected and deployed a high-performance REST API using FastAPI and Docker, wrapping a fine-tuned MobileNetV2 backbone to handle real-time inference with sub-200ms latency.• Integrated MLflow for experiment tracking and model versioning, maintaining a centralized Model Registry to manage transitions between Staging and Production for training iterations.• Designed a Retrieval-Augmented Generation (RAG) style pipeline where classification outputs were fed into a Large Language Model (LLM) via LangChain to generate natural language dietary summaries and predictive health insights.• Developed a robust automated data validation suite to handle noise (corrupt files, EXIF rotation, and out-of-distribution input), reducing production inference errors. • Conducted staged fine-tuning and hyperparameter optimization (GridSearchCV), improving baseline model accuracy while maintaining a lightweight footprint for cost-effective cloud scaling.• Implemented stratified evaluation frameworks, utilizing Macro F1-scoring and per-class precision/recall analysis to mitigate class imbalance and ensure model fairness across visually similar categories. • Engineered a CI/CD pipeline that automated unit testing for model code and integration testing for the API, ensuring zero-downtime deployments using Blue-Green deployment strategies.• Implemented Prometheus and Grafana dashboards to track real-time model health, monitoring for prediction drift and system latency to proactively trigger retraining cycles.

Education

• Raisoni Institute of management

  Master of Business Administration (MBA), Human Resources Management/Personnel Administration, General

  2008 — 2010

• Boston University

  Master's degree, Data Science