← Back Case study · Market intelligence Research pilot · early access

Straat

A market-intelligence desk for the physical AI buildout. The headline names are the peaks; Straat maps what stands behind them: the power, the parts with long lead times, and the companies that depend on each one. Every claim carries its label and its source, every thesis is written so it can be wrong, and every call is scored when it resolves.

Designer and engineer Independent · research pilot 2026 Python · FastAPI · Postgres 17 · React 19

Straat's landscape from above at dawn: the chain runs through the basin as a lit river, stage by stage, with the next wave branching off it — The chain from the mine to the model, rendered live by Straat's own landing pageschematic · no names, no data

00 / 05Context

Beyond the mountain.

Most of the AI buildout sits out of view: in fabs and substations, on launch pads, and in the parts with long lead times that gate all of them. Coverage follows the few names at the top. The constraint that decides the next year usually sits several links further down the chain, in a company nobody is watching.

Research about that chain also tends to evaporate. A thesis lives in a note, the evidence in a spreadsheet, and six months later neither can say why a view was held or whether it survived. Language models make it worse: they will invent a quote for any claim. Straat is built so the record compounds instead, and so nothing reaches it without the sentence it came from.

01 / 05The chain

One chain, from the mine to the model.

The desk places every basket, constraint and company on one order of stages, across four systems: AI, energy, robotics and space. Ore, refining, transport, wafers, lithography, the fab, packaging, chips, networking, power, the halls, the clouds, the models and the demand. Every stage is broken into the parts it is made of, and every edge along the chain carries its source.

The next wave is physical. It leaves the chain after the chips: actuation, sensing, onboard compute, the bodies, deployment. The ontology grows as the desk reads, and amber marks where the chain is tightest.

Over the crest: the contour map rises out of the mist toward the ringed pit where the chain begins — The climb: as the basin opens, the map comes up and the first stage lightsthe landing page's live landscape

02 / 05Evidence

Every line carries its label.

Computed

Reproducible from a named function with named inputs.

Modeled

A formula and its assumptions, both on the line. Never a target.

Interpreted

Read from a filing, with the quote verified word for word before it was stored. A fact the text does not say is refused, and the refusal is kept too.

Speculative

A view without a check yet. Labelled, warned and dated, so a hunch can never pass for a measurement.

03 / 05The desk

Seven layers and one loop.

Data flows up through the desk, and each layer knows only what the one below proved. A fact cannot reach a thesis without its quote; a call cannot reach a surface without its date. Then the record runs back down: the desk proposes, a person disposes, and the record decides what it reads next.

// 01

Sources

Filings, transcripts, prices and open research, each document kept exactly as it was read.

// 02

Quote gate

A fact is stored only with the sentence it came from, matched word for word.

// 03

Ontology

The chain across the four systems, every stage broken into its parts, a source on every edge.

// 04

Theses

Claims, bands and scenarios, each written with what would prove it wrong and how it ends. Superseded, never edited.

// 05

Ledger

Calls with direction, confidence, horizon and date, scored when they resolve.

// 06

Surfaces

What publishes, dated and labelled. Setups and holdings never leave the desk.

// 07

Silicon

The top of the stack: what you read, lit from everything underneath it.

04 / 05The record

Scored, dated, never edited.

Watch

Each thesis names the checks that would break it, and the desk runs them against every new filing and reported figure.

Flag

A failed check raises the breaker. A check that was itself wrong is retracted by a new row that points at the old one; nothing is overwritten.

Score

When a call resolves it gets a Brier score, and the calibration shows what the desk said against what happened.

Ask

Questions about the buildout are answered from the filings, the transcripts and the desk's own record, with the label and the source on every line.

05 / 05Build

What I built.

I designed and built all of it: the desk, its interface, the public site, the film and the design system they share. Scheduled agents read, extract, check and draft overnight; a deep-sleep pass re-reads the week before the first filing lands. Drafts wait for a person, accepted or retired, never merged silently.

The public site is generated from structure alone, and a check that fails the deploy keeps every name and figure from the desk off it. Straat opens as a research pilot: the method is free to read, the daily desk is early access, and no payment is taken.

Desk: Python · FastAPI daemon · Postgres 17 on its own cluster
Interface: React 19 · Vite · one object panel per URL
Models: Local models for extraction · a frontier model for review · no metered keys
Front door: Static · three.js · a landscape from seeded noise
Film: Blender, headless, from the page's own land
Design: STRAAT: one token file for the desk, the site and the brand

Next case study Traffic Emissions →