The last time I dived into building a toy MCP service for a school library assistant, which helps you search for books, borrow copies, and return them. I quickly realized this wasn’t practical. Without authentication and authorization, the service was only useful for an administrator with full access.

What if we want to expose a real service? Security needs to be a first-class citizen.

MCP service itself is not that mysterious. It’s essentially built upon HTTP protocol with an extra layer of protocol and formatting. We can leverage all the concept for building security for HTTP services, like TLS, OAuth, etc. With the FastMCP framework we can extend the Authn/Authz, just like the Python backend development process.

This is a toy implementation, but I think can demo the basic ideas for MCP service and LLM integration.

Code here: https://github.com/hxy9243/agents/tree/main/librarian

See last blog: https://blog.kevinhu.me/2025/08/09/Building-MCP-with-DSPy/

Read More

Recently I’ve been looking into MCP and how it integrates with development framework. I tried with some hands-on tinkering, and learned quite a bit. Here’s my learning and some reflections.

Code here: https://github.com/hxy9243/agents/tree/main/librarian

What’s MCP

Bird’s Eye View

MCP (Model Context Protocol) is a standard way for AI agents to securely connect to your services. It provides a uniform interface for tools, resources, and prompts, enabling seamless integration between AI systems and external services.

Clients like VSCode, Gemini Code, or Claude Code can now integrate with different tools and services and become incredibly powerful, from reading PDFs, search for websites, read your Google Drive files, or book plane tickets for you. You could implement the MCP service once and easily install in all kinds of the AI assistants of your choice.

With this standardization, you can implement your MCP server and they can directly integrate with your user’s agents, assistants, or more complicated workflows. They can call and leverage your tools with a standardized interface.

Think of an AppStore moment for app discovery and AI integration.

Read More

Just realized that I haven’t written down any blogs for about a year. There’s much good learning and experience over the last year that I didn’t get a chance to write it down. It’s a good time to get dust off this blog and get back to the writing habit again.

A Dive Into AI Agent Programming

Recently I got really interested in AI Agent programming. And I spent some time in diving into understanding DSPy. At first I was a little perplexed by starting to read its papers (which focused more on program optimization than actual programming). But once I understood its design I immediately fell in love with it, as its design philosophy resonated with me very strongly.

You can specify and start writing an LLM agent program right away with just a few lines of Python:

1
2
3
4

sentence = "it's a charming and often affecting journey."  # example from the SST-2 dataset.

classify = dspy.Predict('sentence -> sentiment: bool')  # we'll see an example with Literal[] later
classify(sentence=sentence).sentiment

(Examples from https://dspy.ai/learn/programming/signatures/.)

And I can quickly craft a few agents and combine them together to make an agentic workflow.

Read More

1 Introduction

I’ve recently came across a few of Andrej Karpathy‘s video tutorial series on Machine Learning and I found them immensely fun and educational. I highly appreciate his hands-on approach to teaching basic concepts of Machine Learning.

Richard Feynman once famously stated that “What I cannot create, I do not understand.” So here’s my attempt to create a toy implementation of gradient descent, to better understand the core algorithm that powers Deep Learning after learning from by Karpathy’s video tutorial of micrograd:

https://github.com/hxy9243/toygrad

Even though there’s a plethora of books, blogs, and references that explains the gradient descent algorithm, it’s a totally different experience when you get to build it yourself from the scratch. During this course I found there are quite a few knowledge gaps for myself, things that I’ve taken for granted and didn’t really fully understand.

And this blog post is my notes during this experience. Even writing this post helped my understanding in many ways.

Read More

1 Overview

Previously: Paper Readings on LLM Task Performing

I’m still combing through the papers on LLM and I’ll summarize my readings here on this blog. There’s an increasing amount of interest and attention in this field and hence there will be many more papers in the forseeable future. Hopefully I can squeeze more time to read and share my experience with all of you.

Read More

1 Overview

Recently I’ve looked into the LangChain project and I was surprised by how it could be such a powerful and mature a project built in such short span of time. It covers many essential tools for creating your own LLM-driven projects, abstracting cumbersome steps with only a few lines of code.

I like where the project direction is going, and the development team has been proactively including and introducing new ideas of the latest LLM features in the project.

The path to understanding this new project weren’t really smooth. It has its own opinions for code organization and it could be unintuitive to guess how to hack your own projects for more than the tutorials. Many of the tutorials out there explains how to create a small application with LangChain but doesn’t cover how to intuitively comprehend the abstraction and design choices.

Hence I have taken the initiative to document my personal cognitive process throughout this journey. By doing so, I aim to clarify my own understanding while also providing assistance to y’all who are interested in hacking LangChain for fun and profit.

This blog post will dedicate to the overall understanding of all the concepts. I found it really helpful to start by understanding the concepts that directly interacts with the LLM, especially the core API interfaces. Once you have the mindmap of all the LangChain abstractions, it’s much more intuitive to hack and extend your own implementation.

Read More

1 Overview

I’ve spent the last couple of months reading about the development of AI and NLP development in general ever since the release of ChatGPT. And here’s some of my personal findings specifically on task performing capabilities.

The field of AI has been advancing rapidly, and the results have exceeded expectations for many users and researchers. One particularly impressive development is the Language Model (LLM), which has demonstrated a remarkable ability to generate natural, human-like text. Another exciting example is ChatGPT from OpenAI, which has shown impressive task performing and logical reasoning capabilities.

Looking ahead, I am optimistic that LLM will continue to be incredibly effective at performing more complex tasks with the help of plugins, prompt engineering, and some human input/interactions. The potential applications for LLM are vast and promising.

I’ve compiled a list of papers of extending the task performing capabilities in this field. I’m quite enthusiastic and excited about the potential of longer term of this capability that brings LLMs like ChatGPT to more powerful applications.

Here’s my first list of paper, also what I consider to be more fundamental papers, along with my very quick summaries.

Read More

Paper link: https://www.kdd.org/kdd2019/accepted-papers/view/150-successful-machine-learning-models-6-lessons-learned-at-booking.com

Or download.

First published in KDD from booking.com, the paper described its lessons from deploying Machine Learning models in their production service. It provided some intriguing insights. I believe many are very valuable to understanding applying Machine Learning in real-world scenarios.

Here are some of my takeaways.

Read More