← All posts

Why we built atrim.ai

Last week we launched Atrim.ai.

This journey started over a year ago when I first started building using Claude Code during the late summer and early autumn months, filled with half days of salmon fishing and the second half coding and applying lessons I’d learned over the prior 20-odd years of my career focused on application performance management and observability. As the product took shape, I considered not only what was missing but also this new breed of apps and developers who’d shift from hammering out single lines of code to enabling their agents and infrastructure choices to guide the process incrementally.

The engineering effort went largely into building out the scaffolding to support the product as I honed the product direction to be focused primarily on what was needed to make my own job easier with the least cognitive overload while gaining a better understanding of what the agents were capable of, and finding the areas where they were less capable. I certainly had theories that turned into proven facts in some cases and others that turned into wastebin items.

In my mind I kept recalling discussions with my colleague and enterprise product guru, David Pennell, regarding how engineers that are able to reduce code and keep the surface area smallest are often more valuable than ones that can code up hundreds of lines of unmaintainable gobbledygook code. My mission became more clear as I went on - focus only on surfacing the data that is relevant to the human (or agent) needing to solve a discrete problem and prioritize those problems ahead of time.

From my UX background, building on some of the success we had with StreamWeaver, the focus shifted to “least clicks possible” to get to an answer. In some domains you don’t quite know the path to an answer, so you allow the user to wander about, providing enough breadcrumbs to ascertain an answer - but this was the APM of yesterday, associated with SREs and highly skilled individual developers who knew the application by specific lines of code they wrote, which made it easy to transpose into a tool. As with any larger-scale application, and one that is written more (or entirely) by an LLM, this kind of knowledge becomes more esoteric, but what persists is the need to have a precise way to diagnose and fix an issue, ideally the first time and with the highest confidence that the analysis was correct in the first place.

Luckily for me, one of the experts in how to build the kinds of data backends that support large datasets and pinpoint analytics and I started chatting. None other than David, who was itching to get his hands dirty with this new style of agentic development. David brought a high level of discipline and structure to our development processes and has been instrumental in driving the architecture of the platform forward. His experience over 25 years building enterprise-class network and system management tools helped us to streamline our traces, metrics, alerts, logs and incidents ingestion and data pipelines, resulting in a smaller overall product footprint and cost, which we can pass directly to our customers.

We’ve now built a product we’re both very proud of and hope that our customers (and their agents!) will find some delight in using.


Clay Roach is the founder and CEO of atrim.ai. Read more about the team, or connect with Clay on LinkedIn.