Production AI
Pilots do not fail on the model.
The pilot-to-production gap
The demo worked. That was never the hard part.
A great deal of enterprise generative AI stalls between a convincing prototype and a system the business will actually run. The four reasons below are architectural, and they are why a better model does not unblock the launch.
- No data foundation underneath
- A pilot runs on a hand-assembled extract. Production needs governed, current, access-controlled data with a lineage someone can explain - which is a platform problem, not a model problem, and it is usually the real reason the pilot cannot be promoted.
- No evaluation anyone trusts
- A demo is judged by whether it looked right in the room. Production needs a graded baseline, a measurable acceptance bar agreed in advance, and the ability to prove a change made things better rather than merely different.
- No governance surface
- Guardrails, identity and access for agent actions, audit trails, human oversight where it is required. Nobody puts an unbounded agent on live customer data, so this is what actually gates the promotion decision.
- No answer on where the data goes
- Under GDPR the question of where inference runs and what leaves the boundary is not academic. If the answer is uncertain, the security review stalls the launch regardless of how good the model is.
The EU AI Act
Read it as an architecture brief, not a legal one.
Most of what the Act asks for is decided at design time and is expensive to retrofit. Systems built to produce their own evidence satisfy it almost incidentally; systems that were not have to be partly rebuilt. Which of the two you have is an architecture decision taken long before anyone asks.
Classify and document
What the system does, the risk category it lands in, and technical documentation that exists because the build produced it rather than because someone wrote it afterwards.
Oversee and log
Meaningful human oversight where the classification demands it, and logging good enough to reconstruct why the system produced a given output.
Govern the data
Training and grounding data with known provenance, controlled access, and a residency position that holds up to the same follow-up questions the rest of the estate faces.
How the Act classifies your specific system is a question for your counsel and your compliance function. What your architecture can evidence is the question Northkeep answers - the same division of labour as the DORA and NIS2 work.
The engagement
Production AI, designed and built.
Architecture and implementation on Amazon Bedrock: agents, guardrails, agent identity and access, retrieval-augmented generation, evaluation against a graded baseline, and observability that makes failure legible to whoever is on call. Build-versus-buy decided on evidence, and a governance model built to hold up under scrutiny rather than to satisfy a checklist.
Where the foundation underneath is not ready - and it frequently is not - that is said plainly rather than built around. Modernization comes first in that case, because production AI on an estate that cannot carry it is a more expensive failure than waiting.
- Residency for AI workloads
- Where inference runs, what leaves the boundary, and what the grounding data is allowed to touch - designed in, not asserted in a policy.
- Evaluation before promotion
- An acceptance bar agreed in advance and a harness that scores against it repeatably, so "better" is a number rather than an opinion.
- Operable by your team
- The people running it can see why it did what it did. No component depends on the architect still being there.
Evidence
AI that was measured, not just demonstrated.
From our founder's track record, delivered in previous roles and through other firms. The engineering is described in full on the Work page; client figures are published only where cleared.
- Evaluation agreed up front
- An AI search layer on Bedrock and OpenSearch, scored by a repeatable benchmark against acceptance criteria set with the client before the build.
- Inside a live product
- Generative AI built into a B2B software product on Bedrock, with guardrails, agent identity, evaluation, and observability designed in.
- Recommendations at retail scale
- Production recommendation pipelines and serverless APIs behind live customer-facing features for a large multi-brand retailer.
Start here
Modernization and AI Readiness Assessment.
Fixed scope. It establishes whether the foundation can carry production AI, what has to change first, and what the governed target state costs - as a signed evidence pack, not a slide deck. Or skip straight to a conversation about what you want to build.