Sharvani Jadhav

Shipped a production-grade LLM career coach agent using Chainlit and ReAct reasoning, serving 10k users; Integrated NLP-based intent classification, modular tool routing APIs and stateful long-term memory to reduce incorrect responses by 10% and improving personalization across sessions; Architected an end-to-end admissions prediction system for Ivy Leagues with a resume processing pipeline using NER, OCR and GenAI-based feature extraction, enabling automated scoring of 60k applications; Trained and evaluated a 90% accurate ensemble model of XGBoost, Neural Networks and Regression; Deployed a Dockerized production inference pipeline on AWS Kubernetes, processing 15k resumes from S3; Scaled a multilingual data pipeline for skills extraction across 10 languages using DeepL, migrating from pickle-based workflows to PyMySQL to improve query efficiency and optimize storage for 500k records; Engineered a text normalization API to standardize input into entities and metadata, optimizing latency using HNSW-based semantic search in OpenSearch to achieve 100ms retrieval latency across 12k records

Designed a Bayesian traffic-collision model using Negative Binomial regression to handle overdispersed data, optimizing speed-camera placement to reduce fatalities by 40% while cutting installation costs by 50%; Optimized model hyper-parameters via Metropolis-Hastings MCMC using 10,000 samples over posterior

Developed AI ad generator to incorporate brands into videos using Stable Diffusion models and VAEs; Upscaled low-resolution media by 4.5x with improved temporal consistency via hyperparameter tuning