erik.wiffin

Pointwise vs. Pairwise LLM Agent evals and how I used a custom eval tool to speed up my agent by 4x and cut costs by 95%.

LLMs have broken the traditional software playbook: demos are misleading, expertise is shifting too fast to hire for, and the conditions surrounding AI adoption are the worst possible for building the learning organization you actually need.

Mid-way through building an agentic application, I got stuck. I found myself needing to make it work, make it right, and make it fast, all at once, and progress was slow. Reframing it as a distributed system helped me recognize where my problems were coming from and got me going again.

Observability, tool design, and model selection are harder than the orchestration loop. Lessons from building an agentic charting tool with Plan and Execute.

I spent the holidays building an event platform with an AI “vibecoding” tool to see if the results were actually production-ready. While the speed of development was incredible, my audit reveals the critical gaps—from database safety to broken SEO—that still need a human touch before you hit publish.

With large language models capable of generating convincing documents without real thought, this insightful article explores the critical need for new “Proofs of Thought” to ensure genuine intellectual engagement and prevent workplace processes from becoming mere facades.

I have always hated the phrase “Don’t bring me problems, bring me solutions!”. It feels lazy. It incentivizes people to sweep things under the rug if they can’t figure out what to do. It doesn’t give any useful direction to the person bringing you problems, because if they had solutions, they probably would have brought them already! At its best, it’s unhelpful, at its worst, it’s actively harmful. I have similar, but less strong issues with the idea of telling someone to “take more initiative”. That also feels lazy. Particularly when it’s the answer to someone looking for a promotion or generally looking for more responsibility. If one of my reports comes to me asking what they can do to get promoted, they’re taking more initiative right now! They just don’t know what to do.

The demands on a manager’s time are endless and sometimes it feels like you’re being pulled in every direction at once. These demands can make it hard to focus, they make it hard to move from reacting to problems to anticipating them. In this article, I’d like to talk about how I’ve limited demands on my time, counterintuitively doing more with less. The story The very first thing I learned as a new manager was that if I didn’t put a block on my calendar at lunchtime, there was no guarantee that I would eat lunch that day. As an engineer, I thought I knew what it meant to take a lot of meetings, but my promotion brought with it a torrent of people who wanted to talk to me and had no consideration for how hungry I was getting.

As a Director of Engineering, I have monthly 1:1s with all of my direct reports. A 1:1 (one-on-one) is a recurring meeting with no set agenda between a manager and one of their reports. The internet is full of valuable insight into how to run them from my perspective, (ex. The Update, The Vent, and The Disaster) but somewhat more limited in advice on how to make the most of them if you’re on the other side of the table. At Devetry, we run pretty flat, so I have a lot of direct reports, and we only meet once a month. Many people will work more closely with their manager, or meet more frequently, so some of this advice won’t transfer, but hopefully there’s still a few useful nuggets of wisdom.

Or: Why is this software engineer being so difficult? So you’ve got some data. In my case, I’ve been keeping track of how successful my taco truck has been. There’s nothing crazy here: a date column, how many of each kind of taco I sold, and where I parked my taco truck that day. At this point, I’d like to create some sweet visualizations of my data, to better understand the fundamentals of my taco truck business. I know a little python, and I’ve heard it’s good at this kind of thing, so I fire up a jupyter notebook and import it into a pandas dataframe.