Okay, it’s Friday. This is my first LCCTO Code Brain post. Its a little rough, but its a start. I did a tech stack review as part of a startup's due diligence process this week. I am still processing the learnings. It can take me a few weeks to surface some of the patterns I have observed.

The big concept that jumped out this week: 'Mixture of Experts’. It’s a broad term and can be implemented in countless ways. Countless companies used this to describe a phase in the evolution of their AI strategy. An assumption in this term is that not all experts in the mix are equal, some are high quality, and others are rough and experimental. But together they mix to form an evolving and nuanced solution to context and variables AI needs.

As a mix of experts matures, an ‘orchestrating’ expert can be built that understands the role of each sub-expert, and slowly business logic can be assessed and implemented.

New Tech Finds

Building on the ‘mixture of experts’ concept, I dug into how this can be built and used to solve startup challenges. I have known about this concept for a while, but with the focus this last year on Gen AI, the concept has been backburned. Traditionally this concept is implemented at the code level with different ML tools that can be knit together into a black box solution.

But, when I turned on my low-code brain and went looking for ways to get the same effect, I very quickly noticed Make.com. It’s not a perfect solution, but damn it does provide the ability to set up a set of experts.

The Make UI is very very visual. Each expert can be visualized as a set of workflow steps. Each workflow has a name and delivers a specific value. For example, a workflow can process a webhook body from a Zoom call, and extra language signals, terms, and tasks. The same webhook can trigger a number of Make scenarios, each pulling out and contributing specific output.

Insights & Learnings

As I build out a personal ‘mixture of experts’ in Make. I have noticed:

1. 'Expert' is a great mental model for specific agents that solve specific problems.

2. An ‘Expert’ starts simple and gains complexity.

3. Complex experts need to be split into simpler ‘experts’.

4. Simple & targetted agents are easier to support.

Coming up Next…

This coming week I am focusing on finding replacement options for Make.com. I hate ‘first-solution-locking’. This is my word for when I find a good solution and I stop looking. Next week I should have 2 or 3 more equivalents.

Maybe is https://relevanceai.com/ will be the next deep dive.

Finally, I am going deep into Nango.com. This is a low-code solution that can replace several OAuth Token processes and think a code footprint in a codebase. Even more exciting, their pricing page appears to show that a startup can utilize their OAuth Token management services for free. Free is damn sexy.