System diagram
Co-founder project · Hospitality
Cococure Concierge case study
Bookings, availability and recommendations over WhatsApp and phone.
14 days · Agent Pilot
One workflow, automated end to end, with a human approval step and a trace of every run.
For teams moving the same data between the same systems by hand, every day.
Start to finish
14days
- Human approval step
- Every run traced
- CRM, helpdesk, email
The system
Agent Pilot, end to end.
- Already yours
- We build
- Human step
- 01
Trigger
Ticket, email or form
- 02
Agent
LangGraph flow
- 03
Tools
CRM, helpdesk, sheets
- 04
Approval
A person signs off
- 05
Action
Done, with a trace
Use cases
What you get
Six deliverables you can check.
- One workflow automated end to end: support triage, intake or document processing
- Tools wired into your systems: CRM, helpdesk, email, spreadsheets
- A human approval step before anything irreversible
- Tracing of every run: inputs, tool calls, decisions and cost
- An eval set agreed on day 1 and run before launch
- A runbook, and code you own
How it works
14 days, four stages.
01Days 1–2
Map the workflow
02Days 3–8
Agent logic and tools
03Days 9–12
Approval, tracing and failure handling
04Days 13–14
Eval and go-live
What it is, and what it is not.
It is
- One workflow, with a clear before and after
- A person in the loop wherever a mistake is expensive
It is not
- A general-purpose "AI employee"
- Screen-scraping RPA that breaks when a button moves
- Unattended automation of irreversible actions
AI work
Similar systems, already shipped.
Built by Sortup or one of our co-founders, as labelled. Not results of this offer.
Cococure Concierge: Guest (WhatsApp message or phone call), then Concierge agent (PydanticAI on GLM), then Tools (Availability, bookings, recommendations), then Booking (Confirmed back to the guest). - System diagram
Good Boi: Lead (Arrives in GoHighLevel), then Agent (Claude with PydanticAI), then Availability (Read from MoeGo), then Booking (Booked and followed up). Co-founder project · Pet services
Good Boi case study
Booking and follow-up automation across the CRM and the booking system.
- System diagram
Project Parse: Invoices (PDFs in any layout), then Extraction (Fine-tuned Gemini 2.0 Flash), then Review (Corrections captured), then Retraining (A new model from corrected output), then JSON (Fields to your schema). Co-founder project · Document AI
Project Parse case study
Clean JSON from invoices, with a Gemini fine-tuning loop.
The team delivered a complex super app with e-wallet, e-commerce, logistics integrations, and a dynamic commission engine. Their communication was clear, execution was reliable, and they consistently provided solution-driven results throughout the project.
Simon SprouleCo-founder, GatherScope
What changes the scope.
- Systems the workflow touches
- Whether they have usable APIs
- Daily volume and speed per run
- Decision branches
- Approval rules and who signs off
FAQ
Questions about this one.
Which workflow should we start with?
The one with high volume, clear rules and a person doing it by hand today. If you are not sure, the free Teardown picks it with you.
What if the agent gets it wrong?
It will, sometimes. That is why the approval step exists and why every run is traced: a person sees the proposed action and the reasoning before anything happens.
Do we need to change our systems?
Usually not. The agent uses the APIs you already have. Where there is no API, we tell you on day 2 and suggest the simplest route.
Is the Agent Pilot the right first step?
No pitch deck. An engineer, and a straight answer.
- Hours
- We work yours, in the GCC, the UK and EU or the US
- NDA
- We sign yours before the first call
- IP
- You own the code from the first commit