Work

The systems, described honestly.

These are the architectures we build, written out in full. They are patterns rather than client stories — named case studies go up only once a client approves them, and we would rather show you the engineering than a logo wall.

What you are reading

No client names, revenue figures or percentage improvements appear on this page. Nothing here is attributed to a customer, and no metric is claimed that we cannot show you the measurement for.

  • Inbound qualification system

    AI agent
    Challenge
    Enquiries arrive across a form, a shared inbox and a phone line. Whoever is free reads them, guesses at priority, and the rest wait. Nothing is scored the same way twice.
    System
    An agent captures every channel into one intake, enriches the company from public sources, scores fit and intent against written criteria, creates the CRM record, assigns an owner and drafts the first reply for approval.
    What changes
    Sales opens a ranked, enriched list instead of an inbox, and every enquiry is answered — including the ones that used to fall through.
    • LLM agent
    • CRM API
    • Enrichment
    • Queue
    • Audit log

  • Document-to-ledger pipeline

    Document automation
    Challenge
    Vendor invoices arrive as PDFs, scans and phone photographs. Two people retype them into the accounting system, and the month closes late because the backlog builds faster than it clears.
    System
    A processing agent extracts line items, tax and vendor identity from each document, matches it against the purchase order and prior history, posts clean records automatically and opens a review task — with the reason attached — for anything outside tolerance.
    What changes
    The routine majority posts without a human. Finance spends its time on genuine exceptions instead of on data entry.
    • Vision + OCR
    • Extraction
    • Rules engine
    • Accounting API
    • Exception queue

  • Answer engine over internal documents

    Knowledge system
    Challenge
    The answer exists — in a policy PDF, a past thread, a spreadsheet nobody owns. Finding it takes longer than the task, so people guess, and the guesses differ.
    System
    Documents are chunked, embedded and indexed with their source and recency preserved. Retrieval is tuned and evaluated against real questions, and every answer cites the document it came from so it can be checked.
    What changes
    Staff get a cited answer in seconds, and the cases where the system does not know are visible instead of invented.
    • Vector database
    • Embeddings
    • Reranking
    • Citations
    • Evaluation set

  • Operations platform with an AI layer

    Custom software
    Challenge
    The business runs on a spreadsheet that three people edit and one person understands. It has outgrown the tool, but replacing it risks stopping the work.
    System
    A purpose-built application modelled on the real process, migrated in stages so the spreadsheet keeps working until each part is proven. Reporting and anomaly detection are built in rather than bolted on later.
    What changes
    One source of truth with roles, history and reporting — and a system the business owns and can extend.
    • Next.js
    • TypeScript
    • Postgres
    • Background jobs
    • Role-based access
References

Want to speak to someone who has worked with us?

We introduce references directly, on request, once a project is serious enough that it's a fair use of their time. That is a more useful signal than a testimonial we wrote ourselves.

Which of these is closest to your problem?

Tell us, and we'll walk you through what the version built for you would look like.

Start a projectMap a process