talks
I enjoy giving talks on technical subjects, product management, and open source software. I'll travel for it, or record from my home office. Get in touch if your event needs a speaker — or your team needs better technical media.
with Doug Turnbull & Trey Grainger
Jev, Perplexity, Cloudflare's Clef and OpenAI all now sell decision models, which answer typed questions with a calibrated probability. Asked whether each document answers the query, the best of them land next to Cohere, Voyage and Mixedbread untuned, at a fraction of the price. I'll cover what I measured across four models and three request shapes, where they break, and how to pick one.
with Doug Turnbull
A coding agent reads more of your repository in an afternoon than a new hire reads in a month. It writes down every file it opened and every approach it abandoned, and then the session closes and all of it goes in the bin. What survives is usually run counts and token spend, which is enough to bill against and too thin to engineer against. This talk builds the archive instead.
with Doug Turnbull & Trey Grainger
Vendor calculators quote a month you will never have. I price a 100 GB corpus end to end — the writes your index actually creates, every re-index after that, and the build-vs-buy line on one basis — then get to what really decides it: the search system you already run, the credits you already have, and how fast your data grows.
with Doug Turnbull & Trey Grainger
Three patterns that make agentic search work: build digests and indexes offline so the agent retrieves smarter rather than harder, design the result surface for progressive disclosure, and shape the tool loop so the agent knows when it has enough.
with Brian Douglas & John McBride
What an eval actually is, why the experiment phase deserves tooling of its own rather than borrowing the production-monitoring stack, and why downgrading to a smaller model often beats the big one at simply handing off to a tool. Plus MCP adoption, Claude Skills, and the economics underneath AI tooling.
with Steve Tuohy
The observability stack under a distributed database — Prometheus exporter, time series store, Grafana, Alert Manager — and then the operator's side of it: multi-tenant quotas so teams can share a cluster, cross-datacenter replication for failover, and reading a latency dashboard when writes start backing up.
Log Cache put a queryable window over the platform's log and metric stream behind a REST API, so you could ask what just happened without standing up a full logging pipeline in advance to catch it.
Applying SRE practice to the logging subsystem itself: which indicators are worth measuring for log delivery, how to set an objective against them, and what changes about running the system once you have a number you have to defend.
How Loggregator's architecture changed to make log delivery dependable — where the network protocol was losing messages, how buffers handle backpressure, and what service discovery had to do with any of it.
