Pfizer Externship via ExternMaster's student at University of St. Thomas (MN) focused on hands-on data science projects and real-world skills.
Portfolio of projects and practical work done while pursuing my master's, highlighting applied data skills and learning-by-doing.
.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.
Built and orchestrated an AI-agent-driven pipeline (Claude) that generates a full instructional course with interactive, offline-reproducible build-alongs from cached LLM fixtures. Designed a two-stage retrieval pipeline (BM25 vs. LLM-reranked, budget-aware context packing) that improved answer accuracy from 70% to 90% at a fixed token budget; diagnosed a subtle reranking reliability bug and reported a “context rot” null result transparently.
Master's student at University of St. Thomas (MN) focused on hands-on data science projects and real-world skills.
I am Hemanth Shankkar, a graduate Data Science student at the University of St. Thomas (MN). I have some practical work experience and I am pursuing a masters degree while building a portfolio to launch my career in data science.
Externships
Pfizer Advanced: AI-Powered Document Insights & Data Extraction Externship
In progress
Experience
Java Developer
Infosys · Jun 2024 – Aug 2024
Full Stack Engineering Analyst
Accenture · Apr 2021 – Jan 2024
Programmer Analyst
Cognizant · Sep 2019 – Jun 2020
Education
University of St. Thomas, Minnesota
MS in Data Science
PSNA College of Engineering and Technology, India
B.E. in Computer Science & Engineering
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.
The project developed a prototype document-intelligence pipeline that combined OCR, large language models, and retrieval-augmented generation to process enterprise PDFs. The work analyzed LLM architectures and integration patterns, then designed and tested components for text extraction and contextual querying. Deliverables included a research summary on LLM mechanisms and a documented prototype
I researched LLM architectures, evaluated how they process and generate text, and documented the findings that informed the prototype's integration points.
Built and orchestrated an AI-agent-driven pipeline (Claude) that generates a full instructional course with interactive, offline-reproducible build-alongs from cached LLM fixtures. Designed a two-stage retrieval pipeline (BM25 vs. LLM-reranked, budget-aware context packing) that improved answer accuracy from 70% to 90% at a fixed token budget; diagnosed a subtle reranking reliability bug and reported a “context rot” null result transparently.
Building an NLP pipeline that parses clinical text and surfaces medication recommendations and safety flags for expert review, extracting medication entities from clinical notes with Python (spaCy/scispaCy).
Built a recommendation model on a 114K-record dataset with 16 features and an 11:1 class imbalance — applying feature engineering and training/evaluating an SVM (RBF) classifier with Python and scikit-learn to reach 80.5% accuracy and 82% recall.
