How to choose a vector database
Talk abstract. Reading the vector pricing calculators, linking features to outcomes, and what else decides it anyway — existing search systems, AWS credits, and how fast your data grows.
hev mind
Operator notes from Adam Hevenor.
Latest posts
Dope tech: TopK
First entry in a series on the search tech I actually reach for. TopK gets the 2026 architecture right — multi-vector late interaction, object-store-native plus NVMe, LSN read-after-write consistency, and Postgres-compatible SQL. Not enterprise-ready yet, and I say where.
A new role for SWE, managing outcomes not outputs
I haven't reviewed a line of code in over a decade — not because I can't read it, but because I learned in 2008 that trust comes from outcomes, not from meddling in output. That's the same lesson every engineer handing work to an agent is about to learn.
Competing as a privateer in the AI race
Racing is the ultimate expression of engineering, and the privateers — people building their own agent harnesses and competing at the frontier — are where the real innovation in AI is happening in 2026. A field guide, from F1 and downhill MTB to Geoffrey Huntley, OpenClaw, and Pi.
Cheating at search for Astro docs
Doug Turnbull's Cheat at Search course inspired me to build hev ask — agentic Q&A search for Astro docs sites. No vector DB, a committed digest Claude builds offline, and a constrained four-tool-call agent loop at query time.
ask hev: an agentic search box for Astro docs
A ⌘K search overlay that's keyword-instant while you type and turns into a bounded agent when you press Enter. Build-time knowledge graph by Opus, anchor-deep results, Haiku in the loop. Still under development, open source.
New site, who dis
Twenty years of web presence, from a hand-organized folder of files to WordPress to no home at all — and back to a folder of files. This time the files are good.
State of Agentic Engineering: Building a CLI for turbopuffer
turbopuffer wins the price-performance math but left me asking "now what?" every time I loaded data into it, so I built tpuff — namespace browsing, scans, schema management. Porting it from TypeScript to Go, and what a standard harness could and couldn't one-shot in April 2026.
Harnessing agentic search
Search is the fuel for agent context, so solving search for your agent is often the difference between succeeding and failing at a task. Claude Code swapped vector RAG for grep and let the model search for itself; I wanted to try that against Reddit, which is a genuinely gnarly search problem. The result is hiveminer.
The end of vibecheck developer preview, and the start of something more ambitious
I shut the vibecheck developer preview down over the break — I ran out of GCP credits for the database instance. A scatterbrained account of the road there: v0 with my kid, real client work in Cursor after ten years away from code, some painful lab interviews, a DSL built with Claude Code, and what comes next.
Evals were everywhere at re:Invent
Evals were everywhere on the re:Invent expo floor, but the focus felt off — they are becoming observability add-ons rather than a way to reach production. Notes on Anthropic's eval-driven development talk and defensive evals, Datadog's AI tracing booth, Braintrust's mindshare, and why W&B is still the most comprehensive platform for MLE teams.
Open Source Ready Podcast - Rethinking AI Evals with Adam Hevenor
Episode 27 of Open Source Ready with Brian Douglas and John McBride: what an eval actually is, running structured experiments across model variations, and why cost-efficient design matters more than ever — plus MCP adoption, Claude Skills, and the economics of AI tooling. Full transcript included.
Why AI Projects Die in Testing (And How We Fixed It)
A 20-question eval suite on a client MCP integration took nearly eight minutes to run, so every system prompt tweak meant waiting or testing a subset and hoping. Parallel runs, standardized retries, and open-weight models took it under 30 seconds and from $0.35 a run to a penny. That became vibecheck.
Five observations from enterprises using vectors to build AI applications
Vectors are transforming the dynamic AI/ML landscape, ushering in a wave of new use cases while shattering long-standing barriers. But some challenges remain.
Introducing Kube Carbon
Ever wonder what kind of carbon foot print your cloud computing software is producing? What kind of improvement could you have by moving your workload from one zone to another? What about scaling to 0, how does that help? While much of t…
Kubernetes at 4: The worlds next most dominant technology?
This month has marked the 20 year birthday of the search giant Google, and the 10 year birthday of it’s most successful offspring — the mobile operating system Android. Dieter Bohn just published a great piece which lays out how Android…
Five Design Principles for Modern Infrastructure Products
Having worked on Platform as a Service product for the last few years I have come to realize that Product Management and design for cloud products have a few unique aspects to them that make them different than other products. My experie…
Looking Through a Glass Onion of User Feedback
I am working with a new team and Program this month and these thoughts have emerged as a theme among discussions with my Product Management peers. I am also having fun listening to classic albums with my toddler for the first time and em…
Common Misconceptions About Distributed Logging Systems
For many operators logging represents an immutable record of their application behavior and a necessary ledger for auditing, security, and…
Breaking Down the Observability Problem for Distributed Systems
The observability problem space is big enough to deserve a clearer taxonomy and an honest separation of concerns. Starting from "it's just logging" — which is both a useful reality check and the source of most of the confusion — and working outward from there.
Why UX Practitioners Should Learn About SRE
Products expected to just work are hard to research the usual way; nobody offers you feedback on GPS or wifi. Reliability research and SLOs are product design tools, and the perception of poor reliability matters more than poor reliability, which is exactly why UX researchers belong in that conversation.
Rapid Troubleshooting of the Cloud Foundry Logging System
Last year when the Google SRE Team performed their application readiness review of the Cloud Foundry platform they identified two Frequent and Damaging issues related to the log transport system Loggregator. Since then we have developed…
Heuristics For Addressing Product Management Bias — Learning from Human Factors in Avalanche…
Heuristics are rules of thumb, and here’s why they are so effective at managing bias — and some practical examples for Product Managers lifted from the world of alpinism.
Product Thinking : Build What Matters and Ship It!
Product managers should be as comfortable at the command line as they are with sticky notes and 2x2s. What I came to think of as Product Thinking, working alongside a very mixed bench of PMs at Pivotal — some from UX, some from engineering, both effective foundations.
Designing for the Operator Experience
Software has to be easy to operate, scale, and maintain, not just solve the user's problem. Call solving for that Operator Experience. Who operators actually are, and what a year of research with Cloud Foundry operators taught me about designing for them.
Software Architecture for Product Managers
Most PMs learn to read patterns out of user needs, then stall when those patterns meet architecture decisions. The one contribution that matters is asking a frank why — and the uncomfortable part is that plenty of engineers cannot answer it either, because small decisions have surprisingly long life cycles.
10 Steps for Writing Better User Stories
Should designs be attached? Story first, or design first? How detailed should acceptance criteria be? Answering those makes your stories more consistent, but focusing on the life cycle of ideas maps to delivered user value faster. Ten guidelines, from my General Assembly workshop.
Defining Service Level Objectives for Open Source Software
With the publication of Google’s Site Reliability Engineering book many software teams are working to improve their ops practice and start…
A Framework for Understanding Agile Teams
Whatever the methodology, the ability to scale small agile teams within a larger organization is crucial to succeeding in competitive…