Jollen Dai

Built a zero-intervention AI content pipeline using Python at Vaya (YC F24) that autonomously generates travel itineraries, bucket lists, and country guides using Google Gemini, renders Pinterest-optimized images through a multi-tier fallback chain, and auto-publishes via Pinterest API v5 — all to drive organic app installs for the social travel platform. Key engineering: OAuth2 with proactive token refresh, exponential backoff with jitter, Playwright browser automation with CAPTCHA detection as a fallback posting path, 13 Pydantic schema models with validation and auto-truncation, template-driven YAML prompt system, and 20 pytest test modules across the full stack.

• Developed and maintained responsive website with 20+ pages using HTML, CSS, and JavaScript, serving 500+ monthly visitors; managed domain configuration and streamlined content workflows for non-technical stakeholders • Optimized site performance through image compression (AVIF conversion), lazy loading, and CDN implementation, reducing page load times by 40% and achieving 81 Lighthouse performance score • Enhanced search engine visibility through comprehensive SEO implementation including structured data markup, semantic HTML, and optimized meta tags, improving organic traffic by 40%

Designed and implemented comprehensive Python curriculum covering object-oriented programming, data structures, and algorithms for middle school students Achieved the highest participation and retention rates across all school programs through engaging interactive teaching methods that made complex concepts accessible to diverse learners

• Developed multimodal deep learning diagnostic system using TensorFlow and PyTorch with Inception-ResNetV2 and BERT architectures, combining dermoscopic images with patient-specific textual data for improved classification accuracy • Optimized training pipeline using GPU acceleration and pre-computed data augmentation on 40,949 images, enabling efficient transfer learning with Inception-ResNetV2 architecture • Achieved 93.56% classification accuracy across seven skin lesion types with clinical-grade performance (0.9824 AuROC), surpassing published dermatologist benchmarks by 5.2% • Presented first-author research “A Multimodal Approach for Skin Lesion Diagnosis” at Southern California Conference for Undergraduate Research (SCCUR) 2024; submitted first-author paper to Journal of Emerging Investigators (JEI)

• Led development of ML models for Kaggle competitions using Python, scikit-learn, and TensorFlow • Led research analyzing brain tumor scans for brain tumor diagnosis using convolutional neural networks • Organized technical workshops on neural network architectures, data preprocessing, and model evaluation

• Discovered critical 20-40 second safety window where 53% of trespassing events occur before train arrival • Revealed peak risk periods (3-5pm, weekends) by analyzing 3,000+ video clips across multiple locations • Built statistical pipeline identifying location-specific patterns for railway patrol optimization