Why your model regressed at 03:14 AM (and what to do about it)
A walkthrough of three real incidents — and the eval, observability, and rollback patterns that would have caught each of them before a single user noticed.
Read essay
Journal
Field notes from applied AI.
Essays on infrastructure, evaluation, governance, and the unglamorous work that separates demos from products.
Recent
Latest writing.
Speculative decoding without the regrets
How we reduced LLM serving cost by 41% without changing a single model weight — and the three places it bit us in production.
The eval harness we wish we'd built first
An opinionated take on what belongs in your evaluation pipeline before you ship anything to a paying customer.
Introducing Guardrails 2.0
Policy-as-code is now inline at inference — and configurable per tenant, route, and audit class.
On the cost of "good enough" retrievers
A measured look at where investing in retrieval quality pays back — and where it doesn't.
How Northwind cut inference cost 3.1×
A customer engineering story about moving a multi-tenant LLM workload from a generic provider onto Xelvoraa.
The case for boring AI infrastructure
Boring is reliable. Reliable is fast. Fast is what your customers actually wanted.