01 / Current project · Pfizer AI Engineer Externship
Advanced RAG Pipeline
June 2026 – Present
A document-agnostic RAG pipeline over PDFs combining OCR fallback, adaptive chunking, and hybrid retrieval to answer questions over pharmaceutical documents.
- Built a document-agnostic RAG pipeline over PDFs using PyMuPDF extraction with a Tesseract OCR fallback for scanned/image-only pages, ensuring reliable text capture across both digital and non-digital documents.
- Implemented adaptive chunking that auto-switches between whole-page chunks and fixed 512-token overlapping chunks based on document length, preserving table/paragraph integrity on short docs while scaling cleanly to 200+ page reports.
- Designed a hybrid retrieval system combining vector search (HuggingFace all-MiniLM-L6-v2 embeddings) with BM25 keyword search, followed by cross-encoder reranking (ms-marco-MiniLM-L-6-v2) to capture both semantic and exact-match relevance.
- Integrated Google Gemini for answer generation, validating retrieval precision through targeted QA prompts against pharmaceutical documents (storage conditions, QC test methods).
RAGOCRHybrid retrieval
View Advanced RAG Pipeline on GitHub →
📄PyMuPDF
🔎Tesseract OCR
🤗HuggingFace embeddings
🧮BM25
🎯Cross-encoder reranking
✨Google Gemini
02 / Current project
GenAI Bot
June 2026 – Present
A serverless AWS pipeline for answering financial questions over SEC EDGAR data, backed by Claude Sonnet via Amazon Bedrock.
- Architected a serverless AWS pipeline with SAM, Lambda, S3, and EventBridge: deploying a scheduled EDGAR data refresh function and an on-demand document retrieval function backed by Claude Sonnet via Amazon Bedrock.
- Designed Lambda I/O contract with input validation, cross-region Bedrock inference, and structured JSON responses including model metadata and latency tracking.
- Configured Python 3.12 runtime with explicit dependency packaging and IAM-scoped policies, ensuring least-privilege access across S3 and Bedrock resources.
ServerlessAWSGenAI
View GenAI Bot on GitHub →
☁️AWS SAM
⚡Lambda
🪣S3
🔀EventBridge
🐍Python 3.12
✨Claude / Bedrock
03 / Coursework project · ACM@JHU Coding Circles
LaunchStack
Spring 2026
Reverse-engineered and documented a production RAG ingestion pipeline, mapping a 7-stage durable workflow across checkpointed, retry-safe background jobs.
- Reverse-engineered and documented a production RAG ingestion pipeline (Next.js, Postgres/pgvector, Inngest, OpenAI embeddings), mapping a 7-stage durable workflow — route, OCR, chunk, embed, store, finalize, graph enrichment — across checkpointed, retry-safe background jobs.
- Built a failure-mode matrix across OCR routing, normalization, and chunking stages, identifying silent failure surfaces (e.g. unhandled exceptions leaving jobs stuck in a "queued" state, silent fallbacks masking degraded routing) and flagging idempotency risks in non-atomic database writes.
- Diagnosed a status-tracking gap where job state conflated "searchable" with "fully enriched," and proposed an explicit status enum to give downstream teams accurate visibility into pipeline health.
RAGPipeline designSystems analysis
View LaunchStack on GitHub →
▲Next.js
🐘Postgres
🧭pgvector
🤖OpenAI embeddings
⚙️Inngest
