The lab

Before

Monthly reports that take longer to read than to produce.

How

A local Llama model on BrandHub (opens in a new tab)↗'s own server. It is not given the whole database. It has tools to query PostgreSQL and knows which tables hold which data, so it answers in plain language.

Role

I am developing it: the tools it may use, the tables it may read and the checks on its answers.

1. question
2. →Llama selects a query
3. →PostgreSQL
4. →data returned
5. →plain-language answer

Before

About 20 hours to prepare a sprint, and the same routine tasks every week.

How

Claude skills and a chain of agents I built for the most time-consuming parts of my own work: competitor research, git and data-model impact, cleaning up client documents and drafting Jira stories. Atlassian Rovo checks each story for gaps. If your team has work like this, or a portal flow that could be simpler, I would be glad to help.

Sprint preparation fell from about 20 hours to about 6.

Role

I developed them and use them in every sprint.

1. client doc
2. →research agent
3. →impact agent
4. →story drafts
5. →Rovo check
6. →Jira

Before

Currently, a salesperson spends 10 to 15 minutes on a call, takes notes, searches the catalog and builds the order manually.

How

Built for BrandHub (opens in a new tab)↗: the AI captures the order call and drafts the order lines. Calls are in English or Dutch. It reduces administration for salespeople. It does not replace them.

Role

I designed the flow and wrote the specification for BrandHub (opens in a new tab)↗.

1. call
2. →live transcript
3. →model extracts products and quantities
4. →catalog match
5. →draft order
6. →salesperson reviews
7. →quote

Before

A new customer cannot know a catalog of thousands of products.

How

The assistant asks about their business, location, purpose and quantities, then suggests products.

Role

My concept; not yet built.

1. what business?
2. →where?
3. →what for?
4. →how many?
5. →catalog search
6. →suggested basket

Before

Printing machines often produce a few extra units. An order of 1,000 arrives as 1,050, and the supplier invoices for 1,050.

How

A single occurrence is acceptable. The warehouse scans every box, so the system can detect a pattern: +7%, +8%, +6% from the same supplier on the same item. It then alerts product management, who can adjust the default quantity, the tolerance or the discussion with the supplier.

Role

My concept; not yet built.

1. box scanned
2. →matched to order line
3. →difference logged
4. →pattern spotted
5. →alert

Before

About 30 reports a day, and most were questions rather than bugs.

How

The incident assistant from File 03. An LLM checks each report against the system's data models and the code in git, then classifies it as a question it can answer, a genuine bug or an idea for the backlog.

Role

I specified it, wrote the prompt and output rules, and led the launch.

1. report
2. →data models + git
3. →question / bug / idea
4. →answer or Jira

Before

Teams juggle invoicing tools like Teamleader and Moneybird next to their real system.

How

One platform for manual and system-generated invoices, end to end.

Role

I set its OKRs and prioritise its backlog, as part of my four-product portfolio.

1. order or manual
2. →draft
3. →review
4. →booked

Requirements, designs, data models, boards, documentation and chat in one place, so teams no longer move between Jira, Figma, Confluence and Teams. When something changes, AI identifies what else it affects and helps update it.

Role

My idea. I am designing and building it.

Named after the father of pragmatism, a philosophy we follow.

peirce.app →