Last time on Obi the Explorer, I gave the restaurant chatbot a second job. It could take a booking, and now it could answer questions about the menu. A Semantic Router sat near the front and chose which path a message took.
Last time on Obi the Explorer, I wrote about the design of the simplest possible restaurant chatbot. It had one job: to take a table booking. The surprising bit was that the LLM only did two of the seven jobs.
It's still just one Naruto.
Last week on Obi The Explorer, I covered a data pipeline built on Fivetran that turns every Typeform survey into a single reusable set of models, rather than a fresh CSV and a fresh script every time. You can read all about it here. Everything was rosy until a survey came through and exposed weaknesses in the modelling. This post explains why counting people isn't the same as counting answers, what had to change in the models to fix it, and how getting it right ended up mattering well outside the warehouse.
Recently, I've been playing around with cognitive architecture design. Don't worry, I'll dive into it in just a bit. But what I have found is that they make building AI agents easier because you have to actively invest in planning how data flows in and out, as well as the agent's decision-making, reasoning, grounding, and memory.
Our own Groundhog Day.
At my current company, we run patient feedback surveys, clinical intake forms, and stigma studies through Typeform. Think teams across multiple markets and functions. The same way any company interested in providing its customers with the best possible experience would. For the uninitiated, Typeform is a form builder that lets businesses collect data people want to share. It's easy to use, customisable, and it's great. (Not a sponsorship wink wink).
Over the last year, demand for intelligence has skyrocketed. Agentic AI is more popular than ever. However, the popularity of the infrastructure necessary for its acceleration is declining rapidly. Everybody hates datacenters. New York moved to block the construction of new hyperscale data centres, but Starcloud is here to save the day. Or is it?
One file can't hold a rulebook, a process, and the actual grunt work without losing the thread on all three.
That's what our first version of this system did. It doesn't anymore.
Flock Safety's cameras have helped solve over a million crimes. Its audit tool just caught nine cops abusing the system in one day. Both of those facts are true, and neither one settles the actual question: who decides how much surveillance is worth how much safety and who gets to check the checker?
How our data team stopped trusting a coding agent to remember the process.
TL;DR: An AI agent will write you a dbt model that compiles, passes every test, and is wrong. Adding more instructions doesn't fix this, because the real gap is memory, not knowledge. We moved the process out of the agent's head and into the repo: seven phases, each a gate that must be verifiably passed before the next can start. Below is the whole workflow, what triggered it, and what it's caught so far.
Let's have a virtual coffee together!
Want to see if we're a match? Let's have a chat and find out. Schedule a free 30-minute strategy session to discuss your AI challenges and explore how we can work together.