Feb 2025 – Jun 2025 · 4 mos · New York, New York, United States · Remote

• LLM | Audio | Recommendation Systems

Sep 2024 – Dec 2024 · 3 mos · Hoboken, New Jersey, United States · On-site

• Designed assignments, midterm, and final exam papers for over 200 students with detailed rubrics for Prof. Xueqing Liu's course to guide students on common pitfalls and improvement areas.
• Conducted weekly office hours to provide personalized assistance, resulting in a 20% improvement in overall class average post-midterm.

Sep 2024 – Dec 2024 · 3 mos · Hoboken, New Jersey, United States · On-site

• Supported Prof. Jingyi Sun in grading student assignments and conducting research, ensuring accurate evaluation and contributing to academic studies.

May 2024 – Aug 2024 · 3 mos · Chicago, Illinois, United States · Remote

• Led a successful Proof of Concept (POC) to assess the effectiveness and interpretability of prompt engineered versus fine-tuned Large Language Models (LLMs) in extracting multilingual geospatial data to enhance the extraction efficiency of place attributes from text. Accelerated timelines for feature engineering, training, and testing of ML, DL and Small Language Models by 60%.
• Accelerated onboarding of data scientist/engineer by 40% and improved information extraction by 30% by developing an end-to-end AI chatbot with LLMs and Retrieval-Augmented Generation (RAG) on AWS, using FastAPI and ReactJS.
• Containerized and deployed LLMs (Llama-3, OpenChat) on AWS using Docker to detect, extract and format Hours of Operation from website text, optimizing inference efficiency and achieving 85% extraction accuracy.

Apr 2021 – Aug 2023 · 2 yrs 4 mos · Mumbai, Maharashtra, India · On-site

• Expanded HERE Maps global coverage by 17% and saved over $2.5M by leveraging web-crawled data to generate 10M+ high-quality place records. Utilized ML, DL and LLMs to extract key place attributes, including name, category, address, and hours of operation.
• Identified place websites with an accuracy of 92.5% by creating labeled data using heuristics and unsupervised models (K-Means, DBSCAN) for clustering. This data was employed to train Mixture of Experts models (Random Forest, SVM) for classification.
• Extracted street addresses, place names and hours of operation from 9 countries by supervised finetuning foundational models such as T5, GPTJ and DeBERTa on Named Entity Recognition (NER) and Semantic Re-Ranking achieving an overall accuracy of 94.3%.
• Enhanced classification of places across 400+ categories in 6 languages by adapting Transformer(BERT, DeBERTa, XLNet) models to unique regional nuances and improving classification metrics by 7% to 0.88 over previous benchmarks.
• Achieved 25x cost reduction in generating GPS data by building scalable MLOps pipeline on AWS, using optimized CPU / GPU cloud instances and model compression techniques such as ONNX, Quantization and Knowledge Distillation.
• Utilized Prompt Engineering on LLMs to extract place hours of operation and validate model outputs, enhancing data reliability.
• Fine-tuned LLMs like Llama-2 to extract multilingual locale data from website text, achieving 92% extraction accuracy.
• Improved model development by implementing distributed training on multinode Nvidia HPC DGX A100 GPU cluster and automated MLOps pipeline deployment with GitLab CI/CD, Docker, and AWS CloudFormation/SAM.
• Created interactive dashboards using Streamlit and ReactJS for real-time project monitoring, reducing issue detection time by 40% and enhancing stakeholder visibility and decision-making.

Aug 2017 – Apr 2021 · 3 yrs 8 mos · Mumbai Area, India · On-site

• Working as a Deep Learning and Big Data (AI) Engineer, developing Fractal Analytic's AI solutions and capabilities with scaling for structured and unstructured data using Machine Learning and Cloud infrastructure.
• Reported operational risks in client’s critical business functions by developing a scalable ETL solution for processing terabytes of clickstream data with PySpark deployed on AWS EMR, utilizing Jenkins and Oozie for periodic data refresh.
• Worked on Audio Based Recommendation System using Nvidia Rapids and Tensorflow. Used audio features based on the Mel Spectrogram to recommended songs based on Location Sensitive hashing.
• Pretrained and fine-tuned a Speech-to-Text model(Wav2Vec2) on a dataset comprising 16,000 hours of Indian English audio and improved transcription accuracy by 27%. Integrated speaker identification and text-to-speech capabilities.
• Optimized end-to-end data audit, dashboarding and product mapping processes for client sales data by engineering heuristic and machine learning models optimizing delivery timelines by 60% and saving over $500k in operatinal costs.
• Worked on AI Demand Forecasting [CPG] client project using Time Series Forecasting and Analysis along with other ML models. Ensembled and stacked multiple linear, additive, gradient boosting, and time series models. Parallelized multiple existing XGBoost models and reduced training turnaround time by 75%.
• Single-handedly spearheaded the built of a scalable web utility for KPI dashboarding & automation that lets you oversee every aspect of your process in real-time for a [CGP] client.
• Automation client reports using VBA.
• Engineered a selenium solution on company tool reducing workload by 400hrs.
• Cleared AI Engineer Career Track to be designated as Engineer at Fractal Analytics.
• Completed Data Science Boot Camp conducted by Manipal Prolearn.