RAG without vibes: building eval harnesses that clinicians trust
Retrieval-augmented generation is easy to build and hard to make trustworthy. This is how we approached eval for Helio Health's triage assistant.
Engineering notes, case post-mortems and the occasional opinion from the people building these systems.
On 14 August 2025, ETH dropped 34% in 11 minutes. Meridian's clearing protocol processed $380m of positions without a single failed settlement. We're writing up exactly what we built and why it held.
Retrieval-augmented generation is easy to build and hard to make trustworthy. This is how we approached eval for Helio Health's triage assistant.
We reached for Rust on two projects last year. One was the right call. One wasn't. Here's the decision tree we now use.
Every year someone suggests a new database. Every year we end up back at Postgres. This is a love letter and a technical argument.
Glasswork's React Native app has 4.8★ with 2.1 million downloads. The performance work started on day one and never stopped.
We went through SOC 2 Type II with NovaPay. The auditors care about far fewer things than the compliance frameworks suggest — and they care deeply about those things.
Most governance systems are brilliant in theory and unused in practice. We've shipped three. Here's what the users actually engage with.
The question comes up on every AI engagement. After eight production deployments, we have a clearer answer — and it's usually RAG with a twist.
The standard retrospective format is too long, too vague, and produces action items nobody owns. Here's what we do instead.
We've shipped ADRs on every project since 2021. This is the format we settled on and why the two-pager beats the 20-pager.
One email a month. Long reads, case notes, and the occasional technical opinion. No marketing.