the work

Four examples, four very different problems

A month-end that runs itself. A sales team told exactly who to call, and why now. Buying drafted against a 20-year-old back-office system. Certificates verified in the thousands. These are examples, not the full list, and the range is the point: the system a business runs never matters as much as the process inside it.

Every engagement starts the same way: find the workflow that eats the week, then pick the right-sized fix. Sometimes that is wiring AI into what is already there; sometimes it is building new software around it. Each study follows the same shape: the broken workflow, what shipped, the rules it obeys, and an honest status line. Every named client signed off on being named, and each panel is a simulation with synthetic data: the workflow shape is real, the numbers are invented.

01

running monthly

Blu Concepts

Supplying the UK's leading leisure centres

the broken workflow

Blu supply the UK's leading leisure centres with retail stock on sale-or-return: a site only pays for what it sells. It is a process with genuine quirks, which is why off-the-shelf systems never fit and the whole month ran on spreadsheets instead. And Blu did not want software for its own sake: they wanted the admin gone so the team could focus on the people side of the business.

what shipped

  • A system that runs the month itself: it checks in with every site, collects their stock counts and chases the ones that have not answered, without anyone at Blu lifting a finger
  • Reconciliation and invoicing the moment counts land: what sold, what was written off and who owes what, worked out site by site, every invoice drafted ready for the accounts
  • The systems joined up end to end: counts, stock, sales and invoicing in one place instead of scattered spreadsheets
  • A cloud layer on top that the team can simply ask: who has not counted yet, what did a site sell, what is still unpaid, and act on the answers
  • Anything odd held back for a person to check before it goes anywhere

the outcome

A month-end that took days of hand reconciliation now runs itself, end to end: every site checked in with, every count reconciled, every invoice drafted, and a person approving anything unusual before it moves. The time goes back into the relationships with the sites, which is where Blu wanted it.

Time saved will be published once it is measured, not before.

Reconciliation portal · automated invoicing · monthly count cycle with chase emails · Claude queries over live data

  • sale or return
  • monthly stock counts
  • reconciliation
  • automated invoicing
  • chase emails
  • plain-English queries
  • leisure sites
  • AI automation consultant

Simulated example of the monthly cycle the system runs. The count window opens and every site is emailed. Counts arrive site by site; a site that has not answered gets a reminder, then counts. Everything is matched against stock and deliveries, one odd count is held for a person, and the month's invoices are drafted ready for the accounts. At the end the team asks, in plain English, what one site sold this month, and gets the answer straight from live data.

Synthetic data. The workflow shape is real.

every site

checked in with, counted, reconciled and invoiced, monthly

02

running weekly

Taylor Made Designs

UK uniform and merchandise manufacturing

the broken workflow

TMD's sales team used to spend the week working out who to even contact: hunting for companies worth approaching, guessing at the right moment, digging for the right name and email. Meanwhile the back-office system knew plenty and told them none of it. The week went on finding conversations instead of having them.

what shipped

  • A prospecting pipeline that starts from a plain instruction: point it at an industry, or a single company, and it goes and finds the ones worth TMD's time
  • Every candidate weighed the way a good salesperson would: the right size and shape, growing not shrinking, inside its buying window, something happening right now that makes this the moment
  • Then the admin nobody misses: the actual decision-maker named, their email found and verified, anyone already in the CRM filtered out
  • What lands every Monday is a ranked list of exactly who to contact and why, each with a first message drafted in TMD's voice: drafts only, humans send
  • Their back-office system answers questions now too: what have we made today, what has this client bought in the last six months, from a phone, a laptop, wherever the team happens to be

the outcome

The team walks in on Monday to a briefing of who to contact and why now, and spends the week on the relationships instead of the research. Out visiting clients, the answers travel with them.

Priority back-office API · MCP connectors on Cloud Run, OAuth 2.1 · registry and signal enrichment · email verification

  • prospecting
  • buying windows
  • decision makers
  • verified emails
  • CRM dedup
  • drafted outreach
  • ask your ERP
  • AI automation consultant

Simulated example, two scenes. First: the team types a plain instruction, find our next customers in the care sector. The pipeline weighs each candidate the way a salesperson would: is the buying window open, is the company the right size and shape, is something happening right now, who the decision-maker is with their email verified, and whether they are already in the CRM. It lands a ranked briefing of three right-fit companies, each shown with its size and region and an opener drafted; drafts only, humans send. Second scene: out on the road, someone asks what a client bought in the last six months and gets the orders, spend and last-order date on their phone.

Synthetic data. The workflow shape is real.

12

stages between an instruction and a verified name worth calling

03

in production

Nevis Marketing

UK B2B motorcycle-gear distribution, est. 1989

the broken workflow

Nevis run their business on software that is more than 20 years old, and ripping it out was never the answer. £700k to £1m of monthly buying ran from spreadsheets against Pegasus Opera 3, with purchase orders living in SharePoint, invisible to the system of record.

what shipped

  • AI wired in on top: it reads their stock, sales and suppliers, and drafts the weekly buying for a person to approve
  • A purchasing dashboard with a ranked buy list per brand, every line explained
  • Stock health with risk classification, an on-order ledger, and a cash-flow workbench with dual-date payment modelling
  • A data bridge that reads Opera safely, read-only, into a cloud mirror

the rules it obeys

  • AI never invents a quantity
  • Nothing is ever sent to a supplier
  • Nothing writes back to the system of record
  • Every action is audited with before and after state

the outcome

The buyer works from one ranked, explained list instead of parallel spreadsheets, and approves every line. It now sits across £700k to £1m of purchasing a month.

Next.js dashboard · read-only Opera 3 bridge · cloud mirror · full audit trail

  • Pegasus Opera 3
  • purchasing
  • replenishment
  • stock health
  • cash-flow modelling
  • audit trail
  • AI implementation consultant

Simulated example of one line from the weekly buy run. The engine suggests a quantity for a product and the reasoning appears underneath: how many weeks of cover are left, the sales trend, the supplier lead time, and the reorder rule the business set. The engine sizes the number, AI writes the explanation, and nothing moves until a person approves the line. Every approval is logged with before and after state, and nothing is ever sent to a supplier.

Synthetic data. The workflow shape is real.

£700k to £1m

monthly purchasing supported

04

agents in production

CrewPass

Compliance platform for the superyacht industry · Conrad founded it and built its AI engine

the broken workflow

CrewPass is a compliance platform for the superyacht industry. Thousands of crew certificates arrive every month, as photos and scans in every format going, and every one has to be read and checked correctly, because a wrong answer means a crew member does not board.

what shipped

  • Not one agent, a team of them: certificates are identified, read, verified, authenticated and checked for vessel compliance, end to end, the moment they arrive
  • Every check runs in recorded steps, so any decision can be replayed and explained afterwards
  • Changes are tested against a bank of known-good certificates before anything ships
  • Anything the system is not sure about goes to a person, never through on the nod
  • The same approach runs beyond certificates: agents triage bug reports and carry development work, and the team was taught to run them, not just watch

the outcome

Certificate reading runs at better than 95% accuracy in production, up from an 85% baseline, with lookups five times faster and 85% less storage per document. And work that used to take the wider team five hours now ships in about ten minutes, because they were taught to drive the tools themselves.

Compliance is unforgiving. If a certificate is read wrong, a crew member does not board. That is the standard I build to for every client.

Python · LangChain and LangGraph · LangSmith evals · event-sourced pipeline · MongoDB · Google Document AI + Gemini · Cloud Run

  • document verification
  • certificate reading
  • compliance routing
  • human review
  • eval gates
  • audit trail
  • agent teams
  • AI implementation consultant

Simulated view of the certificate queue. Documents arrive as photos, scans and PDFs and move through five recorded steps: identify, read, verify, authenticate and compliance. In this example three certificates verify cleanly in under a minute each, and one with an unclear stamp is not pushed through: it routes to a person instead. The pipeline shape is the real one; the documents and timings are invented.

Synthetic data. The workflow shape is real.

95%+

certificate reading accuracy in production

10 minutes

to ship jobs that took the team five hours, once they learned the tools

Sometimes the first job is the plumbing. Before any AI work, one client needed four businesses consolidated onto one Google Workspace. I did that first.

Bring me the workflow that eats your week

Twenty minutes, free, direct with me. I will ask the right questions and tell you honestly whether it is worth fixing. Prefer to start by email? I reply within 24 hours.