Frank Zeng

Undergraduate Research Assistant @ Indiana University Luddy School of Informatics, Computing, and Engineering

About

I am a Computer Science student at Indiana University (graduating Dec 2025) with a passion for architecting high-performance backend systems and integrating cutting-edge Generative AI into production-grade applications. My expertise lies at the intersection of robust software engineering (FastAPI, Spring Boot) and advanced LLM orchestration (LangGraph, RAG).Key Technical Pillars:AI & Agentic Workflows: I don’t just wrap APIs; I design complex reasoning systems. I developed a production-grade multi-agent stock analysis system using LangGraph, where I implemented a 4-agent parallel pipeline that cut system latency by 75% (from 40s to 10s).Scalable Backend Infrastructure: Experienced in building microservices and RESTful APIs (20+) with a focus on reliability and performance. My projects, such as the AI Financial Platform, maintain 99%+ uptime using FastAPI, MongoDB, and Redis caching mechanisms.Data-Driven Research: My background includes systematic research in 3D Food Recognition under occlusion (PyTorch/CurveNet), demonstrating my ability to tackle ambiguous technical challenges with experimental rigor.Enterprise Internship: During my time at Mobvoi, I optimized SQL query strategies for a 25% speed increase and managed large-scale data pipelines involving 500,000+ hours of speech data.I am a firm believer in "AI-first" engineering—not just as a buzzword, but as a way to solve real-world problems more efficiently. Whether it’s optimizing a vector database (Pinecone/FAISS) or designing a self-reflecting Q&A engine, I strive for code that is clean, scalable, and impactful.Core Tech Stack:Languages: Python, Java, JavaScript, C, SQL.Frameworks: FastAPI, Flask, Spring Boot, React, LangChain, LangGraph, LlamaIndex.AI/Data: RAG, Vector Embeddings (Pinecone), PyTorch, MongoDB, Redis, Docker.Let’s connect! I’m always eager to discuss backend architecture, LLM agents, or full-stack AI development opportunities.

Country

United States

City

Bloomington

Industry

Computer Software

Skill

Retrieval-Augmented Generation (RAG), Software Development, Computer Vision, Deep Learning, Data Science, Python (Programming Language), Machine Learning

Experience

Undergraduate Research Assistant

Indiana University Luddy School of Informatics, Computing, and Engineering

2025-9 - Present · 1 yr 1 mo

Bloomington, IN

I am conducting the first systematic evaluation of CurveNet’s robustness in 3D food recognition, focusing on real-world scanning constraints like occlusion and data sparsity. Key Contributions: Experimental Design: Engineered three distinct occlusion simulation methods (Surface, Random Dropout, Bottom) across 15 scenarios, ranging from 10% to 50% severity, to model realistic sensor limitations. Data Analysis & Insights: Discovered that occlusion location impacts accuracy significantly more than point quantity, with bottom occlusion causing a 64% performance drop—providing critical insights for future 3D perception model architectures. Technical Stack: Leveraged PyTorch and CurveNet to process the MetaFood3D dataset, managing large-scale point cloud data and experimental pipelines. Skills: Computer Vision, 3D Perception, PyTorch, Point Cloud Processing, Experimental Design.

Machine Learning Intern

Mobvoi

2024-6 - 2024-8 · 3 mos

Suzhou, Jiangsu, China

Conducted data collection and preprocessing, developed and optimized speech processing models, and researched advanced methods to improve performance. Collected and cleaned over 500,000 hours of speech and text data, covering multiple languages and Chinese dialects . Employed a custom web crawler and in-house data pipelines for deduplication, anomaly detection, and normalization, reducing manual intervention by 30%. Optimized SQL queries and indexing strategies by using partitioned tables and parallel queries, increasing daily query speed by 25% and reducing resource usage by 10%. Evaluated multiple ASR models across various languages/dialects using diverse datasets and metrics. Leveraged these insights for feature engineering, hyperparameter tuning, and validation, adopting distinct approaches per language/dialect for up to 5% higher accuracy over a single unified model.