Yulin Shi
AI Engineer & Photographer
Building innovative solutions at the intersection of AI and real-world applications.
An AI agent that joins video calls with website visitors, talks to them by voice, and navigates the product website in real-time — clicking, scrolling, and showing features while the visitor watches. Built with a multi-process architecture (Voice Worker via LiveKit, Browser Worker with headless Chrome, API Server) communicating through Redis pub/sub. The agent qualifies visitors, delivers personalized demos, answers questions via RAG, and closes with a meeting or trial offer.
Built a computer vision model for a large-scale pig farm to detect and measure fat thickness from images. Distilled a segmentation model and chained it with a prediction model for end-to-end inference.
A hybrid search engine for African legal documents, deployed in production. Combines vector and keyword retrieval with ElasticSearch 8.9 for accurate results across complex legal corpora. Includes automated metadata extraction for document filtering and management.
An automated pipeline that scrapes daily arXiv papers, indexes them into a vector database, and delivers personalized reports via email. Uses hybrid search to match papers to user-defined topics, with LLM-generated explanations of why each paper is relevant. Fully automated with Celery for scraping, indexing, and report generation.
An image recognition system for vinyl album covers. Uses SIFT for feature extraction and Faiss for efficient indexing and similarity search. Achieves 95% accuracy with 0.03s inference time, optimized for large collections.
Deconstructing algorithms, ethics, and the philosophy of intelligence.
The backbone of this architecture is the choice of Supabase as the Backend-as-a-Service platform, which provides for free not only the relational database but also crucial services such as user authentication and file storage.
How I built Remo, an AI agent that joins video calls, talks to visitors by voice, and navigates a product website in real-time — like an AI sales rep that can actually show you the product.
How I set up an automated research workflow that outperformed my initial solution after a night of autonomous experimentation.
Competing and building under pressure — from AI challenges to live hackathons.
Developed a classification model to analyze data from NASA E-Nose for distinguishing between COVID-positive and COVID-negative breath samples. Applied advanced data preparation and AI techniques to optimize performance despite limited sample size.
Conducted data analysis and preprocessing to ensure quality and relevance. Optimized prediction accuracy through model selection and concatenation and developed a high-performance predictive model to capture market trends.
Developed a model that generates guitar tablature and MIDI files from initial measures, desired style, and key. Using an open-source dataset of guitar recordings, we fine-tuned the Mistral-7B model for coherent musical continuations.
Freezing moments in entropy. A study of light and form.
IMG
DSC
DSC