Tony Yuan

Socratic liveProjects pair you with an AI tutor to help you learn by doing, in real time.

You are an IT engineer at a midsize company where AI adoption is growing rapidly. But as the cloud AI bill climbs and confidential information leaves the company through API calls, management asks you to bring everyday AI work in-house. Across these three liveProjects, you’ll build a private local AI stack from the ground up: research open-weight models, run Google’s Gemma 4 on ordinary hardware with Docker Model Runner, and deploy a ChatGPT-style chatbot for coworkers. You’ll then connect the open-source pi agent to your local model so it can research proprietary documents and write code, before extending your agents with vision, exploring a fine-tuned MedGemma specialist model, and setting up hybrid inference that keeps routine work local while sending more demanding tasks to the cloud. By the end, you’ll have a working on-prem AI setup that runs on a typical laptop with 8–16 GB of memory.

Socratic liveProjects pair you with an AI tutor to help you learn by doing, in real time.

In this liveProject, you’ll take the company’s local AI rollout to the next level by giving its agents the ability to work with images as well as text. You’ll enhance prebuilt agents to process site photos, printed notices, and whiteboard sketches, turning visual information into useful answers and working web pages. You’ll also explore how fine-tuned open-weight models can be adapted for specialized business knowledge, and how local and cloud models can work together in a hybrid setup. By the end, you’ll have a practical path for combining private, specialized local AI with cloud models when additional capabilities are needed.

Socratic liveProjects pair you with an AI tutor to help you learn by doing, in real time.

In this liveProject, you’ll help your company’s local-AI rollout by moving beyond chatbots and putting your private AI model to work as an agent. You’ll connect the open-source coding agent pi to the local Gemma 4 model from the previous project through Docker Model Runner’s inference API. You’ll then use the agent for two practical tasks: researching a folder of company documents and answering questions with citations to its sources, and generating a self-contained web page based on the same documents. Along the way, you’ll see how local models can support autonomous research and coding while keeping company data on-premises.

Socratic liveProjects pair you with an AI tutor to help you learn by doing, in real time.

In this liveProject, you’ll take on the role of an IT engineer helping a midsize company reduce its dependence on costly cloud AI services and keep sensitive data in-house. Your goal is to build a proof-of-concept private alternative to ChatGPT using open-weight models running on local hardware. You’ll survey models and hardware options, set up the cross-platform Docker Model Runner, and run Google’s Gemma 4 E4B model on your own computer. You’ll first interact with the model from the command line, then deploy it through Open WebUI to create a browser-based chatbot that can be demonstrated to management and made available to non-technical colleagues.

Your school district needs you! In this series of liveProjects, you’ll assist them by developing a custom chatbot for teaching students chemistry based on Meta AI’s Llama large language model. To do this, you’ll utilize open-source tools, including WasmEdge, LlamaEdge, Gaia, Git, Qdrant, Python, and UnSloth. Your AI application will need to run with very low resources—it needs to work on the school’s outdated computers—and be equipped with an intuitive user interface. Once you’ve developed your foundation, you’ll supplement it with a chemistry-specific knowledge base to ensure it can’t hallucinate! You will also fine-tune the model to make it work better with agents in the ecosystem while staying safe!