v3 · 2026 · AboutSan Francisco · 2026

Infrastructure
for trust.

We don't sell software. We sell trust. Computer and infrastructure come second.


§ 01 · Thesis

Computer and infrastructure come second.

The enterprises that run the real economy run on documents that look like evidence and decisions that look like judgment. Claims. Invoices. Manifests. Policies. Files that pass through twelve hands before they settle into a record of fact. The software industry has spent forty years asking the world to be tidier. The world declined.

Every Fortune 500 has tried AI by now. None of them trust it to run the work that cannot fail. The gap between a model that can read a document and an institution that will sign on its output is the entire job. That gap is not a model problem. It is a trust problem.

Trust is the product. Determinism is how it's delivered. We build a substrate where each step is inspectable, each output is replayable against the schema that produced it, and each correction flows back into the system that made the call. The math is stochastic. The contract is not.

The compute matters. The model matters. They are infrastructure for the thing the customer actually buys, which is the ability to put their name on an answer and walk into a regulator's office without a folder of caveats. That is the order of operations. Trust first. Everything else underneath.


§ 02 · Careers

Build the trust layer with us.

Open roles below. We hire slowly and we hire people who want to live inside the problem for a long time. If a role looks right, the apply link goes straight to Ashby.

§ 03 · Principles

What we believe.

Five rules we used to write down for ourselves and now write down for everyone. They show up in how we hire, how we ship, and how we hold the line when a customer asks for the easy version of a hard answer.


  1. P-01

    Strong opinions, weakly held.

    Everyone has a stake in the direction. We debate rigorously, bring data to the table, and once a decision is made we execute with conviction. The argument ends at the commit. The work begins.

  2. P-02

    First principles only.

    Old playbooks were written for a world where software demanded a tidy input. That world is gone. We don’t copy patterns because they worked once. We build the context, then solve the problem in front of us.

  3. P-03

    Signal over noise.

    The AI market is saturated with theater. We ignore the demos and the discourse and focus on what shows up in a customer’s P&L. If it doesn’t ship or doesn’t change an outcome, it doesn’t exist.

  4. P-04

    Disagreement nirvana.

    No one is the most right person in the room by default. Intellectual honesty outranks hierarchy. We disagree, we commit, we learn. Then we do it again on the next decision.

  5. P-05

    Act on smells.

    If something looks wrong in the code, in the process, in the strategy, we don’t walk past it. Platforms that hold under load are built by teams that refuse to let small standards slide.

§ 04 · Method

Brick by brick. Function by function.

We don't build an agent. We build a small number of primitives that do one thing each and compose into anything. Each primitive is auditable on its own. Each composition inherits the audit. The platform is the receipt of every step it took to arrive at an answer.

When a customer asks for a new behavior, the question we run is whether the behavior is already a composition of what exists. Usually it is. When it isn't, we add one primitive. Not seven.

F-01
Classify
the gate
F-02
Parse
the surface
F-03
Extract
the substance
F-04
Split
the divisor
F-05
Join
the connector
F-06
Enrich
the context
F-07
Payload Shaping
the contract

Seven primitives. Trillions of compositions. The boring answer to a question the industry keeps trying to make exciting.

§ 05 · TRAJECTORYBEM · SAN FRANCISCOSOC 2 · HIPAA · GDPR
§ 05 · Trajectory

We're going somewhere bigger.

The current product reads documents the rest of the industry refuses to touch and writes structured records the rest of the industry can't verify. That is the foothold, not the summit.

Underneath, we are building a substrate that does not stop at extraction. The same primitives that classify a claim and shape its payload also reason about it, route it, and stand behind the answer when an auditor walks in. We are not announcing what comes next. We are building it, function by function, until the shape becomes obvious.

The companies that run the world will need a layer they trust between the model and the ledger. We intend to be that layer.

Founded
2022
Headquarters
San Francisco
Compliance
SOC 2 · HIPAA · GDPR

§ 06 · Invitation

We are AI skeptics building the next generation of AI.

We are looking for realists. People who would rather build the hard, unglamorous layers that make AI actually work in production than ship another demo. If that's the kind of work you want to do, we want to know you.

BEM · SAN FRANCISCO · CAEST. 2022SOC 2 · HIPAA · GDPR