Quick follow-up to the college search post. Last time, I gave up on fine-tuning a generative model and built a DistilBERT classifier instead. The classifier works great (100% precision, 100% recall), but I always had this nagging feeling that fine-tuning should have worked if I’d found the right model and the right tooling. Today Liquid AI released LFM2.5-350M, and it changes the picture. ...
College search is a 10-dimensional optimization problem that families solve with spreadsheets, gut feelings, and way too many browser tabs. My kid wants to study biology, play D1 soccer, and ideally not pay more than $40K/year. Oh, and not biomedical, just bio. That’s four constraints already. Add in state preferences, school size, acceptance rates, graduation rates, application systems (Common App? ApplyTexas? UC?), and now you’re juggling variables that don’t fit in your head. ...
Originally posted as a thread on X. I was reading that 8090 just landed a partnership with EY last week, and this got me interested in test-driving it. Most of the coverage out there focuses on enterprise usage, I wanted to know if it could help me on an open source project (no budget, PMs, just me/Claude/my repos.) And if this is good for thousands of consultants, maybe it could also be good for the day job. I’ll start by saying that I liked the pitch: the bottleneck in (a lot of, not all obviously) software is deciding what to build, not writing the code. 8090 wants to be the single source of truth that connects product decisions to engineering execution: a gap that’s more or less narrow in various companies, and that doesn’t exist in side projects (where I’m the CEO/CFO/engineer/QA/product marketing person all at once.) ...
This blog has been running Jekyll since 2014. It served me well for years, but the theme was showing its age, the tooling felt stuck, and I wanted something faster and more modern. So I migrated to Hugo. One theme per decade I guess. ...
Originally posted as a thread on X. I’ve been building AI agents for a while now and I’m a bit puzzled about where the industry is going. We’re writing the most autonomous software we’ve ever built, but the tools we use to build them are a mess. Python script calling APIs, state scattered everywhere, YAML configs, framework-of-the-week. It works until it doesn’t, and then you’re debugging at 2am because some MCP gateway is stuck as some tool or API changed shape and nothing caught it. ...
Originally posted as a thread on X. I’m building an agentic harness modeled on Elm’s architecture: a strict functional pipeline where untrusted agent code runs in total isolation. Containers felt like the natural security boundary. Apple’s open-source container tool seemed like the ideal foundation. Building and running containers is dead simple (container build . -t my-image, container run -it my-image), but I still spent a chunk of my weekend (re-)discovering what breaks when you try to get graphical output across a real VM boundary on macOS. ...
Quick thoughts on Sorbet, that has been announced at RubyKaigi 2018 and once again in the 2019 edition, the type-checker for Ruby. ...
Request Routing in the scheduler/container world is an ongoing challenge, with a lot of different and competing solutions that tries to provide a solution. Some solutions are built on top of others, some only support specific schedulers, some operate only at the L7 layer, which doesn’t make things easy for Platform Engineers when it comes to adopting a specific solution. Building and operating Nomad/Consul clusters with the Platform Engineering Team at <<work>> has been an interesting problem to solve. We explored a few different solutions, and as we evolve in a highly-regulated world (we must implement HIPAA, SOC2 Type II, and HiTrust), solutions that weren’t providing basic security (TLS everywhere, poor auditing, etc.) haven’t been considered. ...
I build products regularly, most of them don’t survive their prototype phases. In 2018, I built: a (fast) FaaS with v8 and mruby a cryptocurrency exchange prototype a programmable cryptocurrency trading platform some more stealth projects… (As an aside, my Open Source work isn’t included in this list as I considered it as a “horizontal” supporting these projects, but I’m more and more seduced by the idea of Open Source as a lifestyle business way of living, which I will try to explore in 2019.) I often go as far as setting up Stripe integrations to get the pricing plans in there, but I felt it was too tedious to: Create an account for that new prototype Set up the prices in a spreadsheet, and reflect them there Keeping them aligned with my app’s code and Stripe So in order to automate Stripe’s setup I created a Terraform provider for Stripe. Billing as Code is great! ...
Having spent a lot of time working with the Hashicorp stack lately, I have been working a lot with the HashiCorp stack lately, mostly with Consul, Nomad, and soon Vault. Even if I was more used to operating Kubernetes, I really appreciate the simplicity and focus that HashiCorp builds into its products. I also spend a lot of time in the console (mix of tmux, vim – or Visual Studio Code when pairing with people – and other CLI tools), so I wanted to find a tool that would keep me in the shell, and I eventually released it. Today, I’d like to introduce you to Trek. ...