Currently an Extern @Pfizer
I'm a Senior at UIUC specializing in Computer Science, ready to tackle the tech industry's challenges.
Discover my projects and experiences in Computer Science, showcasing my skills and growth.
.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.
I'm a Senior at UIUC specializing in Computer Science, ready to tackle the tech industry's challenges.
I am a Senior at the University of Illinois Urbana-Champaign pursuing a Bachelor's Degree in Computer Science. I am eager to strengthen my resume and enhance my competitiveness for future roles in tech.
Externships
Pfizer Advanced: AI-Powered Document Insights & Data Extraction Externship
Pfizer
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 externship prototyped AI document-intelligence workflows for pharmaceutical PDFs, combining OCR, image pre-processing, embedding-based retrieval, and RAG experiments. Deliverables included OCR engine comparisons, PyMuPDF extractions with bounding boxes, RAG tests using all-MiniLM-L6-v2 plus reranking, and a blob-segmentation pipeline that labeled pages with doc_type and doc_id.
I extracted 617 words with bounding boxes from multi-page SDFs, compared Tesseract, PaddleOCR, and EasyOCR with recorded confidences and setup notes, implemented hybrid keyword-plus-vector RAG with reranking using all-MiniLM-L6-v2, and built a pipeline that segmented a 10-page blob PDF and assigned doc_type and doc_id per page.
I reviewed a packet of vendor documents, noted it contained mixed file types (spec sheets, certificates, graphs, safety data) and varied date formats. I found it was a collection of differently structured pages that could confuse extraction models and flagged inconsistent date formats.
The module presented messy CSVs and nested JSON. I cleaned and normalized tabular data with Pandas, flattened nested JSON into readable dataframes, and applied denoising, CLAHE, and Otsu thresholding to improve scanned-image clarity for OCR.
The module provided multi-page SDFs. I used PyMuPDF to extract 617 words with bounding boxes, noted table cells and multiword phrases were split, and recommended using regex plus proximity of bounding boxes to label dates.
The assignment compared three OCR engines on a scanned pharmaceutical PDF. I ran Tesseract, PaddleOCR, and EasyOCR, recorded confidence scores and install issues, and produced a written comparison noting PaddleOCR had the highest confidences while Tesseract performed well on clean text. Setup…
Project context involved a short technical PDF. I split the document by page, used the all-MiniLM-L6-v2 embeddings, implemented hybrid keyword+vector retrieval and reranking, and documented test findings about quality-control methods and missing storage-condition details.
I tested three RAG configurations on a sample SDF document, captured retrieved chunks and short answers, and compared confidence and notes. Configuration C used a reranker and produced the most focused result, listing Autoclave, Gamma Irradiation, and Flow Rate Test.
The submission showed a pipeline that segmented a 10-page blob PDF, extracted page-level text, and assigned doc_type and doc_id per page using a constrained LLM prompt. The output listed each page with its doc_id and doc_type (10 entries).
I built a Gradio chat UI, wired it to a RAG backend, and implemented features like an answer-length dropdown, chunk-count status messages, a clear-chat button, and sentence-transformers embeddings. I noted chat-history format fixes and UX tweaks needed.
