Explainers on the tech behind FRED: software architecture, model creation, data decomposition and much more.
20 posts
Standard RAG retrieves passages. Some questions need numbers that do not exist until you compute them. CS-RAG retrieves from three model surfaces over 11.9M spatial records, then computes before the model writes anything.
Asking "where in Northern California" means aggregating millions of spatial records with no filter to narrow them. Pre-aggregating into three hexagon resolutions turns that into three lookups under 2 ms each, and the coarse answer constrains the fine one.
A seven-step geometric binary search search sizes a spatial radius query in 1 to 3 seconds, against 8 to 25 seconds for running the pipeline to find out.
How splitting an agent's tools between database reads and computation over the cached result set halved latency on parameter-change turns.
Precision@k needs documents. With a deterministic simulator as ground truth, accuracy becomes error against simulation across benchmark queries
A deterministic simulator gives you unlimited labels with no measurement noise. That changes what R² means, how you sample, and what counts as validation.
When two thirds of your targets are undefined rather than missing, imputation and sentinel values both fail. A classifier gates the regressor, and its errors are asymmetric.
165 million predictions were baked with fuel at $2.24 a gallon. Changing that does not require retraining, because the simulator's cost function separates cleanly into a rate and a price.
A routing call costs 50 to 200 ms and a procurement analysis needs thousands. Precomputing the detour ratio as a spatial property turns each one into an indexed lookup.
Detour is not a fixed ratio. Short hauls average 2.62 times the straight line, long hauls 1.59, because the first-mile penalty is fixed and gets amortised over distance.
.png)
An agent answering spatial questions cannot call a simulator mid-conversation. The storage layer has to change shape first, and that turns one system into two.
Weighting two objectives into one score is standard. The shape of the frontier is not: 5% weight toward fire buys 118.9% more risk reduction for 20.5% more cost.
A billion 30 m pixels became 2.1 million records at 360 m. That choice was made early, propagates into every downstream model, and cannot be revisited cheaply.
Selecting the cheapest subset that meets a target, where evaluating a candidate is expensive and the pool is unbounded. The buffer is the interesting part, and greedy has a cost that arrives twenty years late.
A 220-line system prompt had grown three separate paragraphs defending one step's ordering. Moving the sequence into four graph edges deleted all three, and cut a pipeline run from four model round trips to one.
A streamed answer arrives as hundreds of deltas. Rendering each one is hundreds of React commits for text nobody has read yet. Collapsing them to one paint per animation frame is four lines, and it moves the hard part somewhere else.
A single-page app serves one empty HTML shell for every route, so 21 documentation pages and every blog post shared the same title and description. Writing real HTML per route at build time is a postbuild script, not a framework migration.
.png)
A CMS fetch that falls back to fixtures is convenient in development and catastrophic in production, where it would have replaced fifteen live posts with two. The fix is not removing the fallback. It is asserting that it did not fire.
Two sites share a design system by copying two files. Nothing enforced that, so a check hashes both against the original and fails when they diverge. It also reports when it cannot see the original, which is the part most checks get wrong.
An agent that draws on a map does not draw on the map. It mutates state, and the view is derived from that state, so the prose and the pixels have different producers. Keeping them describing the same thing is the actual work.