An agreed monthly scope covering updates, monitoring, defect fixes, and a steady stream of small improvements.
Choosing an engagement
Which one fits your situation?
Four common starting points. If you are between two of them, the shorter one is usually the safer first step.
AI proof of concept
Best when
You need to know whether an AI approach is feasible on your data before committing to a build.
Typical scope
One clearly stated question and one dataset
Data assessment and a baseline to compare against
A working prototype, not a production system
Measured accuracy, cost per request, and latency
What you get
A prototype plus a written recommendation: proceed, adjust the approach, or solve it without AI.
Shortest engagement. Time-boxed and scoped to a single question.
Production AI system
Best when
The approach is proven and the feature now needs to be reliable, monitored, and cost-controlled.
Typical scope
Data pipeline, retrieval or model serving, and application integration
Evaluation set and regression checks in the release process
Permissions, audit logging, and human review where needed
Cost, latency, and quality monitoring with alerts
What you get
A deployed feature with measurements, documentation, and an owner for each moving part.
Substantial. Roughly comparable to a custom software project of similar scope.
Custom software project
Best when
A process needs a system built around it, with or without any AI component.
Typical scope
Discovery of the process as it actually works today
Data model, application, and integrations
Automated tests, deployment pipeline, and environments
Handover documentation and team walkthrough
What you get
A working system in production, with the source code and documentation delivered to you.
Scales with process complexity and the number of systems it must connect to.
Existing system improvement
Best when
Software already runs the business but has become slow, fragile, or expensive to change.
Typical scope
Assessment with prioritised technical risks
Staged work that keeps the system available throughout
Test coverage added around the areas being changed
Dependency, runtime, and infrastructure upgrades
What you get
Measurable improvement against agreed targets, and a system your team can safely change again.
Flexible. Often runs alongside a maintenance agreement.
Still unsure? Describe the problem rather than the solution. Part of our job is working out
which of these is the right route — including the case where a smaller change solves it and
no project is needed.
Quality
Tested before it reaches your users
Automated coverage of the business rules and integration points, plus accessibility and performance checks, run as part of the release rather than after it.
Security
Data handling agreed before the first line of code
What data a feature may use, where it is processed, who can reach it, and what is written to logs — decided up front, then reviewed again before release.
Development process
Eight steps, in the order they actually happen
Short increments, working software early, and decisions written down as we go.
01
Discover
We learn how the process works today, who depends on it, and where it costs time or money. This includes the parts people work around rather than the version in the documentation.
02
Define
We agree on scope, success criteria, and what is deliberately out of scope. If part of the problem is better solved without new software, this is where we say so.
03
Design
Data model, architecture, and interface flows are worked out before implementation. Decisions and their trade-offs are written down so they can be revisited later.
04
Prototype
The risky or uncertain parts are built first — an unfamiliar integration, a performance question, or whether a model performs well enough on your data.
05
Build
Development happens in short increments you can review. Working software is available early rather than at the end, so direction can change while it is still cheap.
06
Test
Automated tests cover the business rules and integration points. Accessibility, performance, and security checks run as part of the release process, not after launch.
07
Deploy
Release through an automated pipeline into environments that match. Rollback is defined and tested before the first production deployment, not improvised during an incident.
08
Improve
After launch we watch real usage, errors, cost, and performance. Improvements are prioritised against measurements rather than assumptions.
Planning
Scope, sequence, and what is deliberately out, agreed before the build starts.
Design sprint
The risky parts explored quickly, while changing direction is still cheap.
Delivery
Short increments you can review, with progress and blockers said out loud.
Additional steps in AI projects
AI work carries uncertainty that ordinary software does not, so the process includes
measurement stages before anything is committed to production.
Data assessment
Checking whether the data you hold is complete, consistent, and relevant enough to support the intended result.
Model evaluation
Comparing candidate models on your own examples rather than on published benchmarks.
Prompt evaluation
Testing prompt variants against a fixed example set so changes can be compared instead of guessed at.
Retrieval testing
Measuring whether the right passages are actually retrieved, which usually matters more than the choice of model.
Accuracy measurement
Scoring output against labelled examples and reporting the result with its sample size and known weak spots.
Human review design
Deciding which cases need a person, how they see them, and how their corrections feed back into the system.
Cost and latency analysis
Measuring cost per request and response time under realistic load, then tuning caching, model choice, and prompt size.
Security review
Confirming what data leaves your environment, who can reach each feature, and what is written to logs.
Technology capabilities
Tools we work with
Chosen per project, based on your existing systems and the team that will maintain the result — not on what is currently fashionable.
AI and machine learning
Python
PyTorch
TensorFlow
scikit-learn
Large language model APIs
Embedding models
Vector search
RAG pipelines
Chosen per project after testing on your data, not fixed in advance.
Frontend
React
TypeScript
Astro
Modern CSS
React for application interfaces, Astro where the page is mostly content.
Backend
Python
FastAPI
Node.js
PHP
REST APIs
Selected to fit your existing stack and the team who will maintain it.
Data
PostgreSQL
MySQL
MongoDB
Redis
Vector databases
Relational by default; other stores only where the data shape justifies them.
Cloud and delivery
Cloudflare
AWS
Docker
CI/CD
Monitoring
Deployment that can be repeated, observed, and rolled back.
Product and company names are the trademarks of their respective owners. Listing them here
reflects the tools we work with, not a partnership or endorsement.
Tell us what you are trying to solve
Send a short description of the process, product, or system you have in mind. We will reply with a practical technical direction and the questions we would need answered first.
