CASE STUDY / THEN CAME AI

[00:00] Introducing Alkemy Hi, I'm Devin. I work with Team d²l, and I'm one of the creators of Alkemy. Alkemy is a Python framework that structures machine learning projects end to end, from dataset to deployment. We built it for data scientists, but what surprised us is that Alkemy turned out to give coding agents exactly what they need to do ML work reliably. This means the person driving an ML project doesn't have to be a data scientist or experienced coder anymore. Basic Python skills are helpful, but what matters most is using your domain knowledge to describe the problem clearly, decide what good enough looks like, and flag results that don't make sense. In this video, I'll ask Codex about a project we've been building together and walk you through what it's doing and why it matters. [00:49] Working with Codex and Alkemy As I work with Codex, I'll voice over what it's telling us. If you'd like to see exactly what it's doing, I recommend just watching the video through, then come back, slow down the video while it's working, and then pause once you see the final output. In this project, we're looking at a fraud data set with credit card transactions. It's a public data set, but I've done my best to prevent the agent from knowing I got it from Kaggle, so it can't just go out on the web and figure out what the best solution is automatically. I'm doing a tech demo, and my team is evaluating Alkemy. Introduce us to the project and what we've achieved by working together. How much effort do you think we're saving versus using Codex without Alkemy? I'm sharing my screen and we're doing this live, so keep your response less than five bullets. [01:35] How Alkemy Guides AI Agents What we're seeing is the agent is checking workflows and rules before it decides what to do. This keeps them on track and frees you up from having to provide the same instructions over and over again. With Alkemy, agents are less likely to hallucinate or forget critical details because they're instructed to always journal their progress and record your feedback. A predictable project structure and a set of tools prevent agents from having to improvise as your project grows. The more experiments you run, the more the benefits compound. Let's try something else now. I'm about to walk into a standup, take a look at the project, and put together three bullets I can share. Data scientists need to understand performance, and our product manager needs to relay an update across our organization. [02:29] Using Domain Knowledge to Guide the Model One of my favorite parts of this project was that the agent recommended using historical transaction patterns because they're often useful for predicting fraud. Alkemy gave it the awareness to ask whether the features would actually be available at predict time. Without this, we could have spent days or weeks building a perfect model that can't actually be used. I told the agent, "The features aren't currently available, but we should attempt to prove their value in case it's worth putting on the engineering roadmap." From there, the agent was not only able to build a model that met our performance goals, but it was also able to make the case that a scoped investment was justified. [03:07] Deploying Machine Learning Models In Alkemy, deployments are easy. We give you everything you need to stand up an API or run batch scoring. Launch the API for the deployed model, generate some predictions, and show us the results. The agent launched an API and made a few requests so we can see the model in action. Deployments are easy. The same code that gets packaged for production is the same code that trained your model. Deployments can be run on any platform where Alkemy can be installed. [03:37] From Model to Reliable Deployment Whether you're an experienced data scientist or someone with strong domain knowledge and a problem to solve, Alkemy makes all the difference. It isn't just for building models. It's the difference between a model that lives in a notebook and one that ships reliably every time. That's it for today, guys. If you'd like to know more and take a peek under the hood, click the link in the description to reach out to Team d²l. Please like and subscribe, and I'll see you next time.