Herendil

Herendil does consulting and software for applied AI. We help companies put agents to work on their own operations, using what the company already knows.

Most of that work is building a company brain: one place where what a company knows, decides and does is open to its people and its agents. We start with one workflow, measure it, and keep the cost of running models in view from the first day.

This site is a small set of linked notes. A link opens its note next to the one you are reading, so you can follow a thought as far as it goes. How these notes work.

Book a call30 minutes, by video

What a company brain is

A company brain is a company’s working memory, written down and kept current. It holds what the company knows, from documents and data to the reasons behind old decisions. It also records what the company decides and does, so the next person, or the next agent, can find out why things are the way they are.

Most of it is plumbing. Someone has to decide which sources count, keep them fresh, control who may see what, and show where each answer came from. The model on top is the easiest part to replace.

This site is a very small one.

Agents need context, not prompts

A careful prompt helps for a week or two. An agent stays useful when it can reach the right material at the moment it acts: the current policy, the customer’s history, last quarter’s decision and the reason for it.

When an agent gets something wrong, the cause is usually missing or stale information, and a cleverer prompt only hides it. So we spend most of our effort on what the agent can see, which lives in the company brain, on checking its results, and on knowing when it should ask a person.

Inference is an operating cost

Every call to a model costs money and time, and the bill grows with use instead of headcount. Treat it like rent. Measure it per task, give it a budget, and read the statement every month.

The cheaper path is often a smaller model with better context, a cached answer, or a step that needs no model at all. We price these choices before building and keep watching them after launch, because usage changes. Evaluation tells you whether a cheaper setup still does the job.

Start with one workflow

Pick one piece of work that happens often, has a clear owner, and ends in something you can check. Sorting one kind of support ticket. Drafting a weekly report. Reviewing contracts against a checklist.

Build for that workflow end to end, with real data and the people who do the work today. You will learn what is missing from the company brain, what inference will cost at volume, and where people want to stay involved. Then pick the next one.

Evaluation before scale

Before an agent takes on more work, you need a way to tell whether it does the work well. Write down what a good result looks like, collect real examples, and test against them whenever the model, the instructions or the sources change.

Without tests, every upgrade is a guess. With them, you can try a cheaper model to cut inference costs and see whether quality held. We build the test set in the first week of an engagement, before anything grows.

Where people stay in the loop

An agent should know when to stop and ask. Some steps need a person because the stakes are high, some because the rules were never written down, and some because the person wants to stay close to the work.

We find those points with the people who do the workflow today. Then we design the hand-off so the agent arrives with its context attached, and the reviewer does not have to redo its research.

Software, when it helps

Some problems need a tool more than advice. A connector to a system your agents cannot reach yet. A small internal app for reviewing what an agent produced. A harness that runs your evaluations every night.

We build these when they are the shortest path, keep them small, and hand over the code with its documentation. Most of them end up as part of the company brain.

How an engagement works

A first call to hear what you want to change and whether we can help.
A short discovery. We map one workflow, its sources and the people around it.
A build on real data, with an evaluation set from the first week. Where a tool is missing, we write it.
A hand-over. Your team runs the result, with documentation and the numbers to judge it by.

Some engagements end after discovery with a clear no. That is a useful result too.

Longer writing

Some of these notes will grow into essays. Drafts in progress:

Each one takes an idea from these notes, such as the company brain or inference as a cost, and tests it against a real piece of work.

How these notes work

This site is a tiny company brain. It holds a dozen short notes, each linked to the notes it builds on. Open a link and its note appears next to the last one. At the foot of every note, “Linked from” lists the notes that point to it.

Most of it is plumbing. Someone has to decide which sources count, keep them fresh, control who may see what, and show where each answer came from.
From What a company brain is

The address bar keeps the open notes, so you can share a trail or retrace it with the back button. Hover a link, or press and hold it, to preview the note. Longer writing is on its way. To start over, go back to Herendil.

Book a call

A first call takes 30 minutes, by video. Bring one workflow you would like an agent to handle, or a question about cost or context you have not been able to answer.

No preparation is needed. A rough description of how the work gets done today is enough. We will tell you on the call whether we can help and what the first step would be.

Find a time