From the Team
Deep dives on how we build — AI architecture, mapping, and outdoor technology.
How RidgeText answers dispersed camping questions by text, combining land managers, open forest roads, campgrounds and the camping corridors printed on Forest Service motor vehicle use maps, including the scanned ones.
How we took RidgeText's SMS worker from one Pub/Sub message at a time to four concurrent turns while serializing each sender's messages, by claiming turns from a Postgres table with an advisory-locked query and reducing Pub/Sub to a wake-up notification.
Why a language model's own confidence rating can't be trusted, and how RidgeText now asks three independent checks to identify a photo and reports how well they agree.
How a login code becomes someone else's paycheck, who collects the money along the way, and what any company that sends verification texts can do about it.
We replaced our linear memory with three kinds of memory — semantic recall, permanent facts, and transient notes — and learned a lot about how to be funny.
Piling 15+ complex tools into a single LLM was causing context confusion. Here's why we bypassed top-level routing, embraced the Sub-Agent-as-a-Tool pattern, and built a generic multi-agent backend that scales.
A single always-on Cloud Run instance was costing us $160/month for compute headroom we only needed during LLM tool calls. Here's how we reshaped the system around Pub/Sub and a small dedicated worker — and what the latency numbers actually looked like.
One User. Seven Days. 1,074 AI Images.
How we built a Mapbox-compatible map compositor where the LLM orchestrates layers without ever touching GeoJSON — keeping context windows small and rendering deterministic.
Start with a free plan and text RidgeText from wherever your adventure takes you.