The build pillar
AI software development, to production standards
Applications, agents, and retrieval systems that carry evidence: permissions inherited, decisions logged, evaluations run continuously, and an owner named before go-live.
- Typical entry
- Prototype Sprint, two to three weeks
- Ladder
- Proof, Product, Platform
- Handover
- Code, prompts, evals, runbook, named owner
The situation
The distance between a working demo and a system the business can rely on
The demo answers whether the model can do the task. Production answers a harder set of questions: who is allowed to see this document, what happens when the answer is wrong, how do you know the quality has not drifted since March, and who is on the hook at two in the morning. Most stalled AI programmes are stalled at exactly that boundary.
We build across that boundary. Every system ships with an evaluation harness where a model is in the loop, an audit trail that a reviewer can read without our help, and documentation written for the team that inherits it.
Every build ships with
- Permission inheritance from the source systems, not a flattened index.
- A logged decision trail with inputs, outputs, and the human override.
- An eval harness with a versioned test set and a regression gate.
- Cost and latency instrumentation per workflow, not per token.
- A runbook and a named operational owner inside your organisation.
Engagements
AI software development
Where the systems get designed, built, evaluated, and handed over.
2 to 3 weeks
Prototype Sprint
Two to three weeks from workflow to a working prototype on your own process language and sample data
Proof
6 to 12 weeks
Custom AI Application Development
Production applications with authentication, data boundaries, audit trails, and an evaluation harness
Product
6 to 10 weeks
AI Agent Development
Task-scoped agents with human-in-the-loop checkpoints, tool permissions, and logged decisions
Product
6 to 10 weeks
Enterprise RAG and Knowledge Systems
Retrieval architectures designed to survive audit: source attribution, permission inheritance, and evaluation
Product
3 to 6 weeks
LLM Application Engineering
Model selection, prompt architecture, eval harnesses, observability, and token-cost control
Product
6 to 12 weeks
PoC to Production
A twelve-gate path from a working demonstration to a system operations will accept
Product
6 to 12 weeks
AI Integration for Legacy Systems
Connecting AI into ERP, core banking, mainframe, and system-of-record environments without inheriting their failure modes
Product
3 to 6 weeks
AI Automation and Workflow Redesign
Redesign the process first, then automate the part that survives the redesign
Proof
3 to 5 weeks
Generative AI Consulting
Where generative models earn their place in a workflow, and where they do not
Proof
Engagements
Rapid application delivery
Internal software in days rather than quarters, built under the same governance as everything else.
4 to 8 weeks
AI MVP Development
A first product surface with real users, built on evidence rather than on a roadmap slot
Proof
Days for a prototype, 6 to 12 weeks for a governed production application
Lovable Delivery and Enablement
Building production applications on Lovable, plus the governance, enablement, and rollout structure enterprises need around it
Product
3 to 5 weeks
Enterprise Vibe Coding Governance
The controls that make AI-generated software safe to ship: code review standards, secret handling, data boundaries, publish permissions, and maintainability handover
Platform
One quarter per wave
Build Enablement Programmes
Training non-engineering teams to build their own tools safely, with an intake process and a publishing gate
Platform
The ladder
Proof, Product, Platform
Each rung answers one question and ends in a decision, including the decision to stop. You never buy the next rung before the current one has produced evidence.
What we will not ship
- An agent with write access to a system of record and no human gate.
- A retrieval system that flattens permissions to make the index simpler.
- A model in a regulated decision path with no eval set and no drift monitoring.
- A handover that depends on us remaining on a retainer to keep it running.
FAQ
Questions about the build work
- What counts as production for you?
- Authentication, data boundaries, an audit trail, an eval harness wherever a model is in the loop, documentation, and a named operational owner inside your organisation. A demo behind a login is not production.
- Do we own the code?
- From day one. Code, prompts, evaluation sets, calibration data, and documentation in exportable form, with a handover session and a runbook.
- Can you work with our existing stack?
- Yes. Most of the work sits against ERP, core banking, EMR, MES, or a data warehouse that predates the AI conversation. Integration constraints shape the design rather than being discovered late.
- What if the model is not the hard part?
- It usually is not. Permissions, data quality, and workflow change are the hard parts, and we will tell you when a retrieval problem is really a records-management problem.
Next step
Bring the pilot you already have
The most useful first conversation is about a system that already half-works. We will tell you what stands between it and production.