Joseph Lin

About

Internship experience building an end-to-end medical AI pipeline that extracts and normalizes messy medical PDFs, fine-tunes BioBERT to classify tumor-related statements (including negation/ambiguity handling), and integrates image-model outputs (MedSAM/GroundingDINO) for tumor localization and classification. Also worked on an LLM/RAG system to draft IPO prospectuses using retrieved regulatory/financial sources (including SEC EDGAR) with structured prompts to reduce hallucinations.

Key Strengths

• End-to-end ML pipeline development from messy medical literature to model-ready datasets
• Practical transformer fine-tuning with explicit overfitting controls (freezing layers, regularization)
• Thoughtful dataset design for edge cases (negated statements, ambiguous text; added 'irrelevant' class)
• Metrics-driven evaluation and iteration (loss tracking, validation monitoring, precision/recall/F1)
• Applied RAG + structured prompting to reduce hallucinations in high-stakes financial document generation
• Built text extraction and normalization workflow for heterogeneous PDFs (PyMuPDF, sentence splitting, cleanup)
• Efficient relevance filtering using TF-IDF + cosine similarity with awareness of accuracy tradeoffs

Reference Highlights

• Thoughtful and research-oriented
• Proactively identifies issues and resolutions
• Strong end-to-end ownership
• Improved model classification scores through shipped work
• Delivers reliable, useful technical outputs
• Pragmatic tradeoff decision-making (latency vs depth of validation/cleaning)
• Keeps pipelines reliable with messy real-world data
• Effective full-stack context switching
• Thinks about problems end-to-end
• Balances usability and performance in system decisions
• Efficient use of AI coding tools to speed delivery while maintaining quality
• Delivers work quickly and properly for others to build on

Experience

Education

Skills

Languages