HerendilApplied AI · consulting and softwareHRD-001 · Rev A · Sht 1/1

TitleCompany brain, general arrangement

Dwg no.HRD-001

RevA

ScaleNTS

Sheet1 of 1

Request · 30 minBook a call

HRD-001 · Sheet 1 of 1All directions

Herendil · Applied AI consulting and software

Draw the company brain before you build it.

Herendil helps companies connect what they already know to the people and agents who need it. We design the architecture, build the agents, set up inference and measure the results.

Fig. 1: a company brain. Four kinds of source (documents, tickets, conversations, systems) feed one knowledge layer with an index, a graph and access rules. People and agents read from it and write back. Agents call models through an inference router. An evaluation loop takes traces and test cases, feeds results back into the layer and informs the choice of model.FIG. 1 · COMPANY BRAIN, GENERAL ARRANGEMENTFRONT VIEW · NOT TO SCALE · DIMENSIONS IN ROLES3RD ANGLESCALE NTSn SOURCESDOCUMENTSS-01 · DOCS, WIKISTICKETSS-02 · ISSUES, CASESCONVERSATIONSS-03 · CHAT, EMAILSYSTEMSS-04 · CRM, ERPINGEST11 LAYERKNOWLEDGE LAYERINDEXGRAPHACCESSPERSONPROJECTDECISIONSAME RULES AS THE SOURCE2m READERS & WRITERSPEOPLEP-01 · ASK, DECIDE, REVIEWAGENTSA-01 · TRIAGE, DRAFT, FILEREADWRITEREADWRITE3INFERENCEI-01 · MODEL ROUTERLARGESMALLLOCALCALLTOKENSROUND TRIP4EVALUATIONE-01 · YOUR QUESTIONS, YOUR TASKSTRACESFEEDBACKTEST CASESMODEL CHOICE5
Fig. 1 · General arrangementNotes 1–5 below

1Sources

Start with what is already written down.

Most of what a company knows already exists as documents, tickets, chat threads and records in systems like a CRM. Each source has an owner, a format and its own rules about who may read it.

Herendil’s part · SurveyWe map the sources, their owners and their permissions before anything gets built.

2Knowledge layer

One layer between the sources and everyone who asks.

That layer is the company brain. It indexes content, links it into a graph of people, projects and decisions, and applies the same access rules as the source each item came from.

Herendil’s part · ArchitectureWe design what gets copied and what gets fetched live, how the layer stays current, and what stays out.

3People and agents

People ask. Agents do the work. Both write back.

A person looks up a decision. An agent triages a ticket or drafts the weekly report. Both read from the layer, and the useful ones record what they did, so the next reader starts with more.

Herendil’s part · AgentsWe build agents for specific jobs, with named tools and a human approval step wherever a mistake would be expensive.

4Inference

Choose the model like any other part.

Some tasks need a large hosted model. Many run well on a small one, and some can run next to the data. Choosing per task keeps answers fast and the bill predictable.

Herendil’s part · InferenceWe set up routing, caching and spending limits, so you can swap a model when a better one ships.

5Evaluation

Measure it, or it drifts.

Every change to a model, a prompt or a source changes the answers. An evaluation suite built from your own questions and tasks shows whether a change helped, before your team has to find out.

Herendil’s part · EvaluationWe build the suite with you, run it on every change and feed the results back into the design.

Sequence of work

How an engagement runs

  1. 01

    Survey

    We talk to the people doing the work and map the sources they rely on.

  2. 02

    Drawing

    You get an architecture and a build plan, usable with us or without us.

  3. 03

    Build

    We ship a first agent or retrieval layer to production, with evaluations running from the start.

Reference documents

Notes and essays

Change requestCR-001

Book a call

Thirty minutes on where your team waits for information, and what a company brain could do about it.

01
Describe the work
What your team does, and where it gets stuck.
02
On the call
We sketch the sources, the layer and the readers together.
03
Afterwards
A one-page drawing and a suggested first step, yours to keep.
Book a call Requested by