Vanshika Gupta

Built an LLM based multimodal chatbot with Voice interface for generating insights, charts, and custom dashboards from customer data using natural language queries. Implemented Retrieval Augmented Generation (RAG), leveraging Qdrant vector database and Langroid for knowledge base indexing

Applied prompt engineering techniques and fine-tuning to optimize LLM interactions. Incorporated tool-calling to execute actions

Currently developing a multi-agent orchestration framework integrating agents associated to different products using LangGraph, to build a unified AI assistant

Managed a team of 9 as Product Manager, driving the product lifecycle from ideation to launch, collaborating with cross-functional teams, and aligning solutions with business objectives. Engaged directly with high-value clients (Maruti, Nestle, Mahindra) to develop POCs and secure key deals

Developed core product features and algorithms for ML and combinatorial optimization problems, including vehicle routing with resource constraints, bin packing, and forecasting (PyTorch, Pandas, Google OR-Tools)

Led the architecture and development of a configuration-driven multi tenant backend model for SaaS platform from scratch, reducing development effort for customer POC and onboarding by 10x (Pydantic, MongoDB, Python, FastAPI, AWS Cloud, REST APIs, Docker)

Migrated multiple python scripts to a full-fledged centrally-hosted ML-Optimizer microservice consumed by multiple projects within the company using flask. Used MongoDB and AWS S3 to manage tasks, run configurations and generate output

Developed scheduler based pipeline for ingestion of customer data using Celery and Apache Kafka

Built and deployed scalable modules using Python scripting for automating reporting (ETL) workflow, including big data processing (Hadoop, Teradata), querying database (SQL) and upholding data quality using tests (reduced 20 hours/week)

Built time-series forecasting model like ARIMA, LSTM to predict weekly inventory sales using H2O open source ML Ecosystem

Developed root-cause analysis tool to mitigate inventory deficit risk by correlating 100s of data points and applying specific business domain knowledge to automatically flag the reasons for deficit

Led research for detecting clutters (material changes) on geospatial Mars Subsurface Signal (GPR) data using computer vision based semantic segmentation pipeline (Keras, PyTorch, TensorFlow)

Achieved 70% balanced accuracy with UNet architecture based Deep learning CNN framework on data with 0.0005% imbalance; Built parallel routines for signal denoising, multimodal registration, feature generation and lazy loading to handle large data ingestion

Developed various interactive front-end image editing tools for annotating images for ML applications using Canvas

Wrote back-end feature for batch processing images for their product (NodeJS, Angular)

Employed Spectral Clustering-based unsupervised Dimensionality Reduction to reduce size of satellite image with 400 bands to 30 bands with only 1% loss in multiclass cropland classification accuracy, determined using Support Vector Machines (MATLAB)

Compared various dimension estimation/feature selection approaches like Principal Component Analysis to determine best approach