Reshimu · Crown Heights, Brooklyn

Every agent leaves a trace. We make sure it is the right one.

Reshimu builds AI systems for businesses that need them to behave, governs autonomous agents at the moment they act, and publishes elucidated Chassidus. One house, one method.

Work with usGoverned AI for operatorsAssess · Build · Embed RuntimeFour validators, on npmDeterministic. No LLM in the loop. LibraryElucidated editionsEnglish and Hebrew, free

Work with us

AI that runs the business, and behaves.

We design, build and govern AI systems for operators. Every build ships with the Reshimu runtime underneath it, so the agents acting on your behalf are classified before they act.

Book a scoping call All current work

Assess

Audit the workflows.

Where AI belongs in your operation, and where it does not. A risk-ranked automation map you can hand to anyone.

2 weeks · fixed fee

Build

Custom automations, governed.

Designed and deployed by us. Governed by the runtime from day one. Production systems, not demos.

4–12 weeks · scoped

Embed

Your AI infrastructure team.

Ongoing. We own the stack, the context and the governance layer, and keep it honest as the tools change.

Monthly · limited seats

Build · Construction · New York + BogotáNow running

An operations brain for a $100M+ GC

Multinational general contractor

Monday.com, Procore and Bluebeam wired into one governed context.

Active engagement · ongoing

Ask about this work

Build · Property managementNow running

Property OS

RENTROLL.AI

An AI operating system for a real-estate portfolio.

Active engagement · details on request

Ask about this work

The runtime

Agents act. Reshimu decides if they should.

Four validators sit between the agent and the action. Each one is deterministic: no LLM is asked to grade the call an LLM is about to make. Sub-5ms, auditable, and on npm today.

SHOR · grounding validatorTypeScript

import { classify } from '@reshimu/shor'

const result = classify({
  output:  'Q3 revenue was $4.2M from 47 customers.',
  context: 'Q3 numbers: 47 customers signed up, but revenue was not disclosed.',
})

$ npm install @reshimu/shor

Returns0.4 ms

{
  level:          "PARTIAL",
  score:          0.667,
  flagForReview:  true,
  explanation:    "number '$4.2M' not found in context.",
  entities: [
    { text: "Q3",           found: true  },
    { text: "$4.2M",        found: false },
    { text: "47 customers", found: true  },
  ]
}

Nesher פני נשר

irreversible? held.

Classifies every proposed action GREEN, YELLOW or RED before it runs. Red is blocked and a human is asked.

npm install @reshimu/nesher

Docs GitHub ↗npm ↗

Shor פני שור

grounded? proven.

Extracts every addressable entity from the output and checks it against the context the agent was given. No LLM inside.

npm install @reshimu/shor

Docs GitHub ↗npm ↗

Aryeh פני אריה

in scope? checked.

Catches an agent reaching past its declared boundary before the reach contaminates sibling sessions.

npm install @reshimu/aryeh

Docs npm ↗

Panim Adam פני אדם

gray zone? asked.

Detects the genuinely ambiguous call, files a Beiur report, and halts until a superior agent or a human resolves it.

npm install @reshimu/panim-adam

Docs GitHub ↗npm ↗

Adam KadmonAtzilusBeriahYetzirahAsiyah

Five levels, one direction. Intent descends from its source; each level binds the work more tightly to what was authorized. Atzmut OS and the full docs

Watch it runMerkavah: 184 agents, 21 days, governed and ungoverned side by side↗

Writing · Alignment notes

What we're working out.

Notes on alignment from a tradition that has studied delegation, limitation and the trace that survives it for two hundred and fifty years. Engineering and Chassidus, on the same page.

Engineering

Why we don't trust LLMs to classify the call they're about to make

The most popular agent-safety pattern asks the same model to grade its own action. A second call doesn't fix it.

Depth

The Generative Limit

Tzimtzum as limitation-that-creates. Every tier transition is a receiver-oriented bandwidth choice, and alignment moves inward from the edge.

Chassidus

The Ohr Does Not Need Us to Guard It

On ohr and keili, a letter of the Rebbe on bending, and what it means to apply Chassidus to a machine.

All writing The depth series Glossary

Library · OpenChassidus

תשרי ה'תשכ"ג

Tishrei 5723. Four maamarim of the Rebbe, one argument across a month.

From the shofar's narrow end on Rosh Hashanah to the heir standing in his father's own place on Simchas Torah.

Elucidated bilingual editions of Chassidic discourses: the original beside a translation, the elucidation marked throughout, and a full apparatus. Free to read, download and share. Every edition also ships as a מהדורה מבוארת in Hebrew.

Read this edition The whole library

Shimon Rosenberg

Founder & CEO · Reshimu.ai · Crown Heights, Brooklyn

I grew up inside a tradition that takes seriously the idea that structure is knowable: how a higher order transmits to a lower one, how limitation is what makes reception possible, how intention survives delegation. When I started looking at how AI agents are built and where they fail, I kept seeing the same shapes. Not as metaphor. As correspondence. Reshimu is the company built on that correspondence: consulting that pays for the research, a runtime that comes out of it, and a library that keeps us honest about where it came from.

The full story LinkedIn ↗X ↗

Book a scoping call

Tell us what you're trying to automate.

One call. We'll tell you whether it's an Assess, a Build, or not a fit, and what it would cost. No deck, no discovery theater.

Or email hello@reshimu.ai Here for the runtime? Start with the docs Here for the editions? One email when a new one is published