Currently an Extern @Pfizer
I specialize in AI-powered solutions for document insights and data extraction, transforming enterprise workflows.
Explore my work, including AI-driven document insights and data extraction projects that showcase my technical skills and creative problem-solving.
.jpg)
Pfizer · ✅ Verified by Extern · ⏱️ In progress
Prototype AI-powered document intelligence with Pfizer—using OCR, LLMs, and RAG to automate real enterprise PDF workflows and build a standout portfolio project.
Analyzed 1,000 ride-sharing transactions across 3 demand scenarios (low, medium, peak) and found that just 2 features — rider-driver ratio and time-of-day — explained 80% of fare variance; this insight gave the operations team a simple, data-backed rule for when to apply surge pricing rather than relying on gut instinct. Tested 3 pricing models (Linear Regression, Random Forest, XGBoost) and identified XGBoost as optimal; the resulting model reduced pricing error by 24% over baseline, meaning fewer underpriced rides during peak demand and less overcharging during low demand — directly improving revenue yield and customer fairness simultaneously.
Analyzed 14,640 tweets about 6 US airlines from February 2015 and found that 63% of all customer sentiment was negative — identified that United Airlines and US Airways together accounted for 46% of all complaints, with Customer Service Issues being the single largest driver of negative feedback at 20% of all negative tweets. Built and deployed a full-stack NLP application: trained 5 classification models (Logistic Regression, Naive Bayes, SVM, LSTM, BERT) with BERT achieving highest accuracy; built a live Streamlit web app with a real-time tweet sentiment classifier and interactive airline insights dashboard — deployed to Streamlit Cloud with a public URL, making findings immediately accessible to any stakeholder without technical setup. Translated findings into an actionable airline operations brief: fixing the top 3 negative reason categories could convert an estimated 40%+ of negative sentiment — directly quantifying the customer experience impact of operational improvements.
Designed and assessed scikit-learn regression models to predict ski resort ticket prices based on operational features including vertical drop and night–skiing capacity. Applied EDA and SelectKBest feature selection to identify key predictors and reduce multicollinearity, cutting the feature set by 35% while maintaining predictive accuracy.
I specialize in AI-powered solutions for document insights and data extraction, transforming enterprise workflows.
I am an emerging professional exploring new career paths, currently focused on applying AI technologies to real-world challenges. Through my externship at Pfizer, I am honing my skills in document intelligence and data extraction, utilizing cutting-edge tools to drive innovation.
Externships
Pfizer Advanced: AI-Powered Document Insights & Data Extraction Externship
Pfizer
Experience
Data Scientist Fellow
Springboard · May 2025 – Present
Assistant Professor (Research Cell)
Poornima University, Jaipur · Aug 2023 – Jan 2025
Research Fellow
Department of Physics, MNIT Jaipur · Jan 2018 – Dec 2022
Research Intern
CSIR–National Physical Laboratory, New Delhi · Jan 2016 – Aug 2016
Education
Malaviya National Institute of Technology, Jaipur
Ph.D. in Physics (Material Science) · Class of 2023
Punjabi University Patiala, Punjab
Masters of Science in Physics · Class of 2016
Chaudhary Devilal University Sirsa
Bachelors of Science · Class of 2014
Skills
✅ Verified by Extern · ⏱️ In progress
Prototype AI-powered document intelligence with Pfizer—using OCR, LLMs, and RAG to automate real enterprise PDF workflows and build a standout portfolio project.
I developed an AI-powered document intelligence prototype for Pfizer, focusing on automating PDF workflows using OCR, LLMs, and RAG technologies. This project enhanced my understanding of AI applications in the pharmaceutical industry and allowed me to create a meaningful portfolio piece demonstrating innovative data extraction methods.
I mapped document types and extraction challenges from a 30-page vendor bundle and produced preprocessing routines that denoised and enhanced scanned images to improve OCR reliability.
I reviewed a vendor document package, noted three date formats, identified the oldest document (Oct 2022) near the 3–4 year review window, and listed OCR, layout, and boilerplate-text challenges that would hinder automated extraction.
The work processed a heavily noisy scanned image through denoising, CLAHE contrast enhancement, adaptive thresholding, and morphological opening/closing. Iterative parameter tuning reduced static and produced a clean binary image ready for OCR.
I processed multi-page SDF PDFs, used PyMuPDF to extract word-level bounding boxes, grouped words into lines and columns, applied regex and anchor‑proximity rules to locate dates and vendor fields, and produced structured field outputs and notes on extraction challenges.
A scanned pharmaceutical supplier PDF contained rotated timestamps, small italics, and signatures that Tesseract misread. I pivoted from label matching to content-aware extraction, described preprocessing steps, and showed OCR errors and their fixes (examples included misread tokens and rotation…
During the Build a Simple Chatbot task I debugged deprecated Gemini model calls and Colab Secrets misconfiguration, updated model and authentication, and ran a Python chat loop that successfully sent prompts and received model responses.
Analyzed 1,000 ride-sharing transactions across 3 demand scenarios (low, medium, peak) and found that just 2 features — rider-driver ratio and time-of-day — explained 80% of fare variance; this insight gave the operations team a simple, data-backed rule for when to apply surge pricing rather than relying on gut instinct. Tested 3 pricing models (Linear Regression, Random Forest, XGBoost) and identified XGBoost as optimal; the resulting model reduced pricing error by 24% over baseline, meaning fewer underpriced rides during peak demand and less overcharging during low demand — directly improving revenue yield and customer fairness simultaneously.
Analyzed 14,640 tweets about 6 US airlines from February 2015 and found that 63% of all customer sentiment was negative — identified that United Airlines and US Airways together accounted for 46% of all complaints, with Customer Service Issues being the single largest driver of negative feedback at 20% of all negative tweets. Built and deployed a full-stack NLP application: trained 5 classification models (Logistic Regression, Naive Bayes, SVM, LSTM, BERT) with BERT achieving highest accuracy; built a live Streamlit web app with a real-time tweet sentiment classifier and interactive airline insights dashboard — deployed to Streamlit Cloud with a public URL, making findings immediately accessible to any stakeholder without technical setup. Translated findings into an actionable airline operations brief: fixing the top 3 negative reason categories could convert an estimated 40%+ of negative sentiment — directly quantifying the customer experience impact of operational improvements.
Designed and assessed scikit-learn regression models to predict ski resort ticket prices based on operational features including vertical drop and night–skiing capacity. Applied EDA and SelectKBest feature selection to identify key predictors and reduce multicollinearity, cutting the feature set by 35% while maintaining predictive accuracy.