Skip to content

See what we build.

Three systems we built ourselves, with the reasoning left in. They are demo implementations, not client engagements, and none of the situations below describes a paying client.

Client work is covered by agreements with the agencies we build for. When a partner is happy to be named, real case studies will replace these.

Demo implementation

AI lead routing

A new lead is enriched, qualified, assigned to the correct salesperson and followed up automatically.

The situation

A marketing agency receives leads from a client’s Webflow site across three service lines. Sales owners are assigned by hand each morning, so the first reply goes out the next working day and the reason a lead was prioritised is never written down anywhere.

Implementation decisions

The model returns a schema, not prose
Qualification asks for a fixed JSON shape: service line, fit band, and a one-sentence reason. Anything that fails validation retries once, then routes to a human queue instead of guessing.
Routing rules live outside the prompt
Owner assignment is a lookup table the agency can edit. Changing who covers a territory does not mean re-testing a prompt.
The reason is written to the CRM
The qualification sentence is stored on the deal record, so a salesperson can see why a lead landed with them and disagree with it.
AI lead routingDemo workflow
  1. Webflow form
  2. Enrich company
  3. AI qualification
  4. Route by fit
  5. HubSpot deal
  6. Slack to owner
IDLEStatic view6 steps

Built with

  • Webflow
  • n8n
  • OpenAI
  • HubSpot
  • Slack

What it changes

Leads are assigned and acknowledged without a morning triage pass. Measurable results depend on the client’s own volume, so no figures are claimed here.

Demo implementation

Intelligent document processing

Incoming documents are classified, extracted, stored correctly and converted into structured actions.

The situation

An operations team receives supplier documents as email attachments. Someone opens each one, decides what it is, renames it, files it in Drive and copies a handful of fields into a spreadsheet.

Implementation decisions

Classification and extraction are separate steps
One call decides the document type, a second extracts against the schema for that type. Splitting them keeps each prompt short and makes a wrong answer easy to locate.
Low confidence stops the workflow
Extractions below the threshold are held in a review queue with the source file attached. Nothing partially-read reaches the database.
The original file is never modified
The source attachment is stored untouched alongside the extracted record, so an extraction can be re-run later against a better prompt.
Intelligent document processingDemo workflow
  1. Gmail attachment
  2. Classify type
  3. Extract fields
  4. Confidence check
  5. File to Drive
  6. Write record
IDLEStatic view6 steps

Built with

  • Gmail
  • n8n
  • Anthropic
  • Google Drive
  • PostgreSQL

What it changes

Documents arrive filed and structured. The review queue makes the failure rate visible instead of hiding it inside a spreadsheet.

Demo implementation

Automated client follow-up

Follow-up logic reacts to lead status, booking activity and previous interactions.

The situation

Follow-up is a person remembering. A lead books a call and still gets the nudge sequence; a lead who replied twice gets the same message as one who never opened anything.

Implementation decisions

Booking cancels the sequence, immediately
A calendar event matching the contact stops every pending step. This is checked at send time, not only at enrolment, because bookings happen mid-sequence.
Channel follows the last real interaction
If the contact last replied on WhatsApp, follow-up goes to WhatsApp. Consent state is checked before either channel is used.
Quiet hours are enforced in the contact’s timezone
Messages queue rather than send outside working hours. The workflow holds them instead of dropping them.
Automated client follow-upDemo workflow
  1. Stage change
  2. Check calendar
  3. Check replies
  4. Choose channel
  5. WhatsApp or email
  6. Log to CRM
IDLEStatic view6 steps

Built with

  • HubSpot
  • n8n
  • WhatsApp
  • Google Calendar
  • Email

What it changes

Follow-up reflects what the contact actually did. The agency can see every message the system sent, in the CRM timeline.

Have a client asking for something you can’t easily build?

Send us the requirement. We’ll tell you how we’d approach it, what it would require and whether it makes sense to build.

Discuss a client project

No pitch deck required. A rough description is enough.

Or write to [email protected]