Haoran (Casper) Zhang
Software Engineer · AI/ML Systems
I build LLM systems end to end — data pipelines, evaluation frameworks, and user-facing AI products.
Recently: LLM extraction pipelines over biomedical literature, multimodal benchmarks for large vision-language models, and the GPU serving stack underneath.
M.S. from Yale University · B.S. in Computer Science, University of Washington - Seattle
Projects
Ruby's VibeCourt
AI mediation system · Agentic workflow · Local-first web application
An AI-guided mediation platform that interviews each side separately, transforms unstructured narratives into structured case intelligence, and produces a transparent, actionable resolution brief.
Adaptive private interviews · Structured evidence and timeline extraction · Comparative conflict analysis
Next.js · TypeScript · Tailwind CSS · Dexie / IndexedDB · LLM Agents
FitNutri AI
RAG health analytics · Full-stack web app · Personal project
A retrieval-augmented nutrition assistant that answers health and food questions using research literature and USDA nutrition data.
- Built a RAG pipeline over 5K+ research abstracts - Improved answer relevancy by 73% through systematic evaluation - Delivered 94% accuracy on structured USDA nutrition lookups
Python · LangChain · Llama 3.1 · ChromaDB · Django · React
WhisperNote Agent
Local-first AI transcription tool · CLI · Personal project
A privacy-first tool that converts local audio and video into timestamped transcripts, summaries, and structured analysis notes.
- GPU-accelerated transcription with automatic CUDA-to-CPU fallback - Modular AI provider layer supporting local, manual, and API workflows - Reproducible sessions with speaker segments, timestamps, and metadata
Python · faster-whisper · CTranslate2 · Typer · OpenAI SDK · CUDA
Experience
Yale School of Medicine
NLP Research Intern — NIH-Funded PFAS Water Project
Research · Jun 2025 – Feb 2026 · New Haven, CT
Built an LLM pipeline that mines biomedical literature at scale to extract structured data on PFAS water contamination.
Screened 18.7M PubMed papers down to 3,611 relevant studies, then extracted 39,061 entities across 7 types with schema-constrained GPT-4.1 prompts.
Reached 0.94 F1 (98% precision / 90% recall) via an error-driven prompt-tuning loop; wrote annotation guidelines and managed 6 annotators (IAA 0.75).
LLM extraction · Prompt engineering · Neo4j
Yale University
LVLM Research Assistant — Multimodal LLM Benchmark
Research · Dec 2024 – Jun 2025 · New Haven, CT
Built a benchmarking framework and serving stack for large vision-language models on ophthalmic imaging.
Benchmarked 5 SOTA LVLMs across 16+ clinical tasks; contributed novel eye-related multimodal datasets and ETL pipelines (+32% data, +40% rare-condition coverage).
Integrated vLLM and tensor parallelism for 100B+ parameter models on 4×A100 GPUs — 8.2× speedup and 65% lower VRAM.
LVLM · vLLM · Benchmarking
University of Washington
ML Research Assistant — Low-Resource Environmental Monitoring
Research · Jan 2023 – Dec 2023 · Seattle, WA
Designed efficient CNN-Transformer models for on-device fine-grained image classification.
Built a hybrid CNN-Transformer MoE (sparse experts + MobileViTV2) with 4× parameter reduction (4.4M → 1.08M) while keeping competitive accuracy.
Used clustering-based expert routing (K-means on patch embeddings) to remove auxiliary losses; cut FLOPs 8% with under 1% accuracy drop.
Efficient ML · MoE · Computer Vision
JICHUANG Technology Co., Ltd.
Data Scientist Intern
Industry · Jun 2023 – Sep 2023 · Taiyuan, China
Built recommendation and prediction systems for an education platform.
Built a hybrid course recommender (collaborative filtering + BERT content embeddings), improving top-5 precision by ~20% and reducing cold-start.
Trained Random Forest / XGBoost models for dropout-risk prediction at 90% accuracy (+15% over baseline), enabling early intervention.
RecSys · XGBoost · NLP
Selected Publication
LMOD+: A Comprehensive Multimodal Dataset and Benchmark for Developing and Evaluating Multimodal Large Language Models in Ophthalmology
Published · ACM Transactions on Computing for Healthcare · 2026
A multimodal dataset and benchmark evaluating large vision-language models on ophthalmic imaging tasks.
Contact
Open to Software, AI/ML, and Infrastructure Roles — 2026
Yale M.S. · building AI products independently · open to full-time roles. Reach me by email.