Reference builds

How we think about the problems we take on.

Written as engineering briefs rather than client stories. We do not name clients or publish numbers we cannot evidence — when a project is cleared for publication with measured results, it will appear here as a full case study.

AI & Automation

Multi-system operations pipeline

The problem

A team receives work through email and forms, re-keys it into two internal systems, and tracks status in a shared spreadsheet. The process is slow, the spreadsheet drifts out of date, and nobody can answer where a given request is without asking someone.

The approach

Put a single intake in front of every channel and make the pipeline the source of truth. Incoming requests are normalised into one schema, validated, enriched where a model can reliably extract structure from unstructured text, then written to the downstream systems through their APIs. Anything ambiguous is queued for a person rather than guessed.

What gets delivered

Unified intake across email, web forms and API
Validation and structured extraction before anything is written
Exception queue for cases needing human judgement
Full execution log — every run, its inputs and its result

Outcome

Re-keying disappears, status is answerable from one place, and failures surface as alerts instead of being discovered weeks later.

Stack

Python
FastAPI
n8n
PostgreSQL
Docker

Mobile Engineering

Field application with offline operation

The problem

Staff working away from reliable connectivity need to record work as it happens. A connection-dependent application means either paper and later re-entry, or lost records.

The approach

Treat the device as the primary store and the server as eventual. Writes go to a local database immediately and queue for sync; the interface never blocks on the network. Reconciliation is explicit about conflicts rather than silently choosing a winner, and anything unresolvable is surfaced to the user instead of dropped.

What gets delivered

Single codebase deployed to both iOS and Android
Local-first writes with a durable outbound queue
Conflict handling with a visible resolution path
Media capture with deferred, resumable upload

Outcome

Work is recorded once, at the point it happens, and reaches the back office without a second data-entry step.

Stack

Flutter
Dart
SQLite
Node.js
Firebase

Web Platforms & SaaS

Multi-tenant platform and admin console

The problem

A growing product keeps customer data in one shared schema, handles permissions in application code, and requires an engineer to run a query whenever the support team needs an answer.

The approach

Enforce tenancy at the data layer rather than trusting every query to remember it, so a missed filter cannot leak across customers. Move permissions into a role model checked in one place. Then give support and operations their own console, so the questions that used to reach engineering are answered by the people who have them.

What gets delivered

Tenant isolation enforced below the application layer
Centralised role and permission model
Operations console for support and account management
Versioned REST API with generated documentation

Outcome

Support answers its own questions, engineering stops running ad-hoc queries, and adding a customer is configuration rather than a migration.

Stack

Next.js
React
Python
FastAPI
PostgreSQL

Recognise your problem in one of these?

The specifics will differ, but the shape usually does not. Tell us what you are dealing with and we will tell you how we would approach it.

Start a project What we do