# The best systems feel invisible.

URL: https://levystreet.com/systems/
Updated: 2026-10-19

How Levy Street designs, builds and runs production AI: embedded with operators, built around your systems and measured against an operating number.

## Short answer

Levy Street builds production AI in three stages. Discover: map the workflow, constraints, handoffs, data and approvals with the people who do the work. Prove: build the narrowest production-grade system that can prove value against an operating metric. Scale: add monitoring, evaluation, governance and support so the system keeps improving in use. Every system is built around existing software and measured against a number the business already tracks.

- Embedded with the work: engineers sit with operators, not just stakeholders.
- Built around your systems: permissions, data, approvals and existing software.
- Measured in production: cycle time, volume, conversion, cost to serve or revenue.
- Lab agnostic: the model is chosen by testing, not preference.

## Embedded with the work

Most AI projects are scoped in meeting rooms. Ours start where the work happens. We sit with the people closest to the workflow, collect real examples (emails, forms, reports and the awkward edge cases) and map the constraints that actually shape the job. That is where the narrow, valuable opportunities are found, and where the risks show up early.

*[Image: Three stacked data planes with a glowing network on the middle layer]*

## Built around your systems

Models are only useful when they are connected to permissions, data, tools, approvals, reporting and the software your team already depends on. We design every system around your systems of record and operating rules, not the other way round. See our [integration guides](/integrations/).

## Measured in production

Every deployment is instrumented against an operating metric: cycle time, support volume, conversion, cost to serve, throughput or revenue. If a system cannot be measured, it cannot be trusted or improved.

## Discover, Prove, Scale


1. **Discover.** Map the workflow, constraints, handoffs, data and approvals that define the real operating problem. Choose one workflow and one number.
2. **Prove.** Build the narrowest production-grade version that can prove value: real data, real users, evaluations, guardrails and observability from the start.
3. **Scale.** Harden integrations, document ownership, train the team, connect monitoring, evaluation and governance, and generalise the patterns that become lasting capability.


## What we bring

| Capability | What it means in practice |
|---|---|
| Agentic workflow design | Deciding where an agent helps and where a fixed workflow is safer. See [automation versus AI agents](/compare/automation-vs-ai-agents/) |
| Evaluation harnesses | Test sets built from your real cases, run on every change. See [LLM evaluation](/services/llm-evaluation/) |
| Tool and API integration | Least-privilege access to the systems the work touches |
| Human approval loops | Fast, logged review for anything that changes records or reaches customers |
| Security and permissions | Prompt injection testing, data boundaries and audit logs. See [AI security reviews](/services/ai-security-review/) |
| Model selection | Choosing models by testing on your workflow. See [AI model selection](/services/ai-model-selection/) |
| Production observability | Quality, latency and cost tracked in production. See [monitoring AI quality](/engineering/monitoring-ai-quality/) |
| Team enablement | Training so your people can run, question and improve the system |

## Lab agnostic by design

OpenAI, Anthropic, Google and the open-weight ecosystem all have different strengths, and the line-up changes every few months. We test them against your workflow rather than defaulting to a vendor. When a better or cheaper model arrives, a well-built system can switch with an evaluation run, not a rebuild. Our guide to [AI model migration](/engineering/ai-model-migration/) explains how.

## Questions people ask

### Why do so many AI pilots fail to reach production?

Usually because the hard part was never the model. It is data access, permissions, exception handling, review paths, monitoring and the judgement of the people who know the business. Pilots that skip these look good in a demo and stall when they meet real work. We build them in from the first version.

### Which AI models do you use?

Whichever does the job reliably at the right cost: models from OpenAI, Anthropic, Google and the open-weight ecosystem, plus conventional software where it is better. We test candidates against your own examples before choosing.

### Who owns the system when you are finished?

You do. We document how it works, train the people who run it, hand over the code, prompts, evaluations and runbooks, and agree what support, if any, you want from us afterwards.


Published by Levy Street (https://levystreet.com), Wellington, New Zealand.