Werk 12 — who carries it
A model is the smallest replaceable part of an AI product. Everything that decides whether it can be shipped, changed or trusted is engineering: where the call is made from, what context it is given, what it is allowed to do with the answer, how a regression is caught before a user meets it, and what a change costs in money and milliseconds.
This werk builds that layer. It assumes the model will be swapped — for a cheaper one, a newer one, a self-hosted one — and makes that swap a routine release rather than a rewrite.
The offer
An AI system that can be operated by an engineering team rather than babysat by its authors: one governed path for every model call, an evaluation gate in CI, and a cost and latency budget that holds under load.
A single governed chokepoint for every model call — routing, retries, quotas, PII masking and lineage — with a CI gate that fails the build on any call made outside it.
Chunking, embedding, indexing and retrieval built against the real corpus, with the retrieval quality measured rather than assumed.
Golden sets, rubrics and offline plus online evaluation wired into the pipeline, so a prompt or model change ships on evidence.
Token and inference budgets per route, caching and model routing, with the fallbacks that keep the product answering when a provider does not.
Gateway, retrieval and evaluation audit with a remediation sequence.
Reference architecture, provider arbitration, release gates.
The AI platform built, gated and operated by the client team.
Used in these offers
State of the proof
Experiences that prove it
Sub-capabilities and evidence
Publishable figures
Questions
- How is this different from the AI werk?
- The AI werk decides which model belongs behind which decision and proves it is worth putting there. AI engineering builds the system that call lives in: the gateway, the retrieval path, the evaluation gate, the cost budget and the release mechanics. One chooses the intelligence, the other makes it operable by a team that did not write it.
- Why insist on a single model gateway?
- Because governance is only enforceable at a chokepoint. Masking, routing, quotas and lineage are each trivial to implement once and impossible to enforce across scattered call sites, so the gateway is the difference between a policy and a claim.
- What state is the proof in?
- The data-platform and lifecycle lines are proven in client mandates — BNP Paribas FLOA and AXA — and the conversational lines at Orange Business and Peetchr. The gateway, retrieval and evaluation lines currently live inside the On-Kare venture rather than a client case, so they read HELD rather than PROVEN.