A tool that fits. Checks you can inspect.
Build focused internal software and evaluate AI against the work it needs to do. Start with a concrete task and a result your team can review.
Explore the work behind the interface.
Client software contributions and evaluation work show the approach. Each linked project states its actual scope.
Choose an example to watch the task become a useful result.
See the week without combining spreadsheets.
A small dashboard shows finished jobs, open quotes and missing invoices in one place.
- Problem
Every Monday someone copies job sheets, invoices and open quotes into one spreadsheet.
- What goes in
Exports from the job system and the accounting tool.
- What comes out
A weekly view the owner checks before anything is sent.
The full workflow, step by step
The complete workflow
Connect exports
Read the agreed exports in their current format.
Match records
Match jobs, quotes and invoices by reference number.
Build the view
Show finished, open and unmatched items.
Verify
Test totals against the source files and flag mismatches.
What each tool does
- Claude Code
- Assists implementation within a reviewed development workflow.
- Codex
- Assists coding, investigation and verification.
Where it runs and where data goes
Runs where your files already live: a workstation, a company server or an agreed hosted service. Access rules are part of the build.
Your team stays in control
The owner checks the totals and decides what follows up.
When something goes wrong
An unmatched invoice is listed as unmatched, never guessed. A missing export stops the report with a clear message.
Give one awkward process a useful interface.
A focused tool keeps its inputs, status and next action together.
- Problem
A recurring task moves between messages, files and separate spreadsheets.
- What goes in
Your current process, sample records and access rules.
- What comes out
A small application or integration tailored to the agreed task.
The full workflow, step by step
The complete workflow
Map the task
Follow one real request through the current process.
Build the flow
Create only the views and integrations the task needs.
Test the edges
Check missing data, permissions and failed connections.
Hand over
Document operation, ownership and update responsibilities.
What each tool does
- Claude Code
- Assists implementation within a reviewed development workflow.
- Codex
- Assists coding, investigation and verification.
Where it runs and where data goes
Deployment is chosen around your users, existing systems and data requirements. Local, hosted and hybrid options are scoped explicitly.
Your team stays in control
Your team reviews the working version before release.
When something goes wrong
Failed connections need a visible status and recovery path, rather than silently losing the request.
Test the answer. Check how it is graded.
A task, reference answer and verifier make the checks inspectable.
- Problem
A plausible answer passes review without a repeatable test of its correctness.
- What goes in
The task, source material, expected result and explicit scoring criteria.
- What comes out
An evaluation package with tests, findings and documented limits.
The full workflow, step by step
The complete workflow
Define task
Specify a realistic task and the evidence it requires.
Write reference
Build an inspectable solution and scoring rules.
Challenge grader
Test correct, empty and deliberately incomplete submissions.
Report scope
Explain which behaviours the evaluation covers.
What each tool does
- Python / SQL / Docker
- Implements repeatable checks and an isolated verifier.
Where it runs and where data goes
Evaluation data and execution environments are agreed in advance. Synthetic examples are identified separately from client material.
Your team stays in control
A person defines the scoring criteria and reviews what the tests do not cover.
When something goes wrong
An empty or unsupported answer must not receive credit because of a weak grader.
What we deliver. What we need from you.
You receive
- A working version of the agreed tool or evaluation package.
- Relevant tests, setup documentation and an explicit handover.
- A defined deployment, update and support arrangement.
Your team provides
- The problem, representative data and users for feedback.
- Existing system access and the permissions the tool must respect.
- Acceptance criteria and a person authorised to approve release.
Give the next build a clear purpose.
A short conversation to confirm the task, the connections and the pilot scope. No purchase is required to start the discussion.
Discuss this setup