#74 AI Engineering at Typeform, New Podcast & LLMs in the Browser

DanielJuly 21, 2026July 23, 2026

It’s Daniel here, the author of 📘 Building Al Agents with LangGraph.js and currently working on 📙 Understanding Neural Networks from Scratch. 

Before we start, what was an interesting thing you discovered lately? Hit reply and share it with me. I’m always looking for new ideas, tools, and resources to explore.

Let’s see what’s new:

· Recorded a new podcast with Faris Aziz, Staff Engineer at Smallpdf and organizer of ZurichJS. We talked about engineering fundamentals, the story behind ZurichJS, career lessons, what moved the needle for him & what turned out to be a time waster.

· Speaking of ZurichJS, the conference that will take place on the 10th and 11th of Sept 2026, Faris was kind enough to leave a 15% discount for the main conference ticket: JSCRAFT. Speakers include names such as Dominik Dorfmeister (TanStack Query), Scott Tolinski (Syntax.fm), and Kevin Powell.

· In the previous edition, we took a deep dive into the state of the software developer job market. This chart from Levels.fyi serves as a perfect map to put things into perspective.

· Susanna Wong, Staff AI Engineer at Typeform, was featured in a great interview on the Seniors at Scale podcast. Big shoutout to Dan Neciu for producing such a good podcast! There were many good points about how the Typeform team manages the integration of AI-powered features into their product. I’ll list some of my main takeaways in the below notes.

· Under the hood, the Typeform AI Agent is made up of a supervisor agent and sub-agents for specialized tasks. The orchestration is done via LangChain and the agents communicate with each other via A2A. The internal Typeform APIs are exposed as tools via MCP. At the moment everything is powered by an OpenAI LLM.

· Evals are the “biggest elephant in the room for any team building AI features”. Each update to the prompts, adding a new version of the underlying LLM, or changes in the configuration may lead to unexpected results.

· Evals require a mindset change compared to deterministic features. You don’t run the eval once, it passes, and move on. Instead, you have to run it quite a few times, compare the different results, and get a general sense of whether your current setup and agent will be able to handle situations like this. It’s not 100%, but it gives you confidence that your prompt changes and configuration will handle that type of user request.

· The team at Typeform uses offline and online evals. With offline evals, you set a predefined set of evals, run them during development, and gain confidence that things can work. Offline evals can be run locally, on CI/CD, or in cloud environments. With online evals, once you deploy your agent, you can track real user requests and model responses as traces. From those traces you can see whether performance has drifted or improved against actual user requests so you can continuously observe and evolve.

· Web AI allows the browser to come with locally integrated LLMs. The local models are already down from a couple of GB to 100 MB. Using them gives you privacy, they don’t charge per token, they work offline, and they don’t need an external API call, therefore they are much faster.

· Local LLMs are great tools for fundamental features like detecting profanity and translations. For deep reasoning and complex tasks, you can still use the frontier models. Think of it like having some local browser storage versus making a call to an Oracle database.

· If Web AI becomes a thing, then most likely Chrome will ship with a Gemma model, Edge with some OpenAI alternative, and Safari with Apple Intelligence. Will this take us back to the pre-jQuery era when we’ll have to deal with browser compatibility? But this time with regard to local LLM capabilities instead of CSS rounded corners?

· Susanna also gave this really interesting talk titled A Front-end dev’s hobbyist approach into Machine Learning a while ago about the fundamentals of neural networks. The reason I like it is that, alongside really good explanations, it also comes from the perspective of JS-focused developer.

· And speaking about the fundamentals, my book 📙 Understanding Neural Network – From Writing if else Code to Training AI Models is 50% done. Currently I’m reviewing the draft of the first part. Remember that you can join the waiting list and get a 20% discount.

And that’s all folks for this edition. Summer is passing very fast so Make sure you enjoy it while it’s here.

Be kind, be smart, and keep coding!