State of the proof
What this is not
It is not a chatbot practice, and it is not model training for its own sake. The work starts from a decision a business already makes and asks whether a model improves it against a baseline — if there is no baseline, the first deliverable is the baseline.
It is not applied to an estate that has not been made readable. Data engineering comes first: an estate nobody can query cannot be automated, and applying models to it produces confident answers from unreliable inputs, which is worse than no answer.
It is not an autonomy pitch. An agent is scoped by the tools it can reach and the failure it is allowed to have, and where that boundary cannot be drawn the honest recommendation is a narrower system that a human still closes.
Applied AI
A model put behind a decision a business already makes, with the baseline that says whether it improved anything.
Applied AI puts a model behind a decision a business already makes, and measures whether the decision got better. That framing does the filtering: a decision nobody currently makes has no baseline, no owner and no way of telling success from a plausible-looking output.
The first deliverable is usually the baseline, not the model. What does the current process get right, how often, and at what cost — because a model that is eighty per cent accurate is excellent against a coin flip and a regression against a rule nobody had written down. Skipping this step is the commonest way an AI programme becomes unfalsifiable.
From there the work is ordinary engineering with an unusual failure mode. Serving, latency and cost behave like any other system; correctness does not, because the system returns something confident whether or not it knows. So the scope includes what happens when it is wrong: the fallback, the human check where one is warranted, and the monitoring that catches drift rather than only outage.
It is bought when there is a decision at volume, a record of how it has been made, and an owner for the outcome. It is the wrong purchase when the data is not yet readable — that is data engineering first — or when the real requirement is a report, which a query answers more cheaply and more honestly than a model.
An AI feature is a product decision wearing a model. The engineering that decides whether it survives contact with production is upstream and downstream of the model: the data it is fed, the evaluation that gates it, and the trace it leaves.
Phoeniks venture: an APSIM-DSSAT-ORYZA model ensemble for agricultural yield prediction.
Retrieval over a knowledge graph built from a 15,500-file engineering corpus.
On-Kare: model versioning, training-data lineage, incident cases and explainability specified as platform-level aggregates.
Agentic systems
Systems that act rather than answer: the tool boundary, the evaluation harness and the failure mode you accept before it runs unattended.
Evaluation is the part that decides whether a system ships. A harness that scores the agent on the cases that actually occur, including the ones where the right answer is to stop and hand back, is built before autonomy is widened. Without it, the only available signal is whether a demo went well, and a demo is not a distribution.
Agentic work starts where the demo stops. A model that answers well in a chat window is a capability; a system that books the appointment, files the claim or moves the record is an operator, and it inherits every obligation an operator has — a scope, a stop, a trace and a review.
The collective's own delivery runs on a supervised multi-agent pipeline; no client case published yet.
Model Context Protocol servers and tool contracts wired across the collective's build stack.
Knowledge-graph retrieval built over a 15,500-file codebase corpus on the On-Kare venture.
On-Kare: every model call routed through one governed provider, enforced by a hard CI gate with no exemption list.
On-Kare: injection-pattern detection, PII masked before the model call, fail-closed consent guards.
Evaluation dimensions, rubrics and reference datasets specified per AI phase before implementation.
On-Kare: model, prompt hash and timestamp logged on every AI output for conformance audit.
On-Kare risk tiers T0 to T3: the lower the tier, the harder the stop — no auto-dismiss, dual control, hard blocks.
How it is bought
Any expertise can be bought in six forms: fixed price, time and materials, short placement, long placement, permanent placement or named seat. Each specialty carries the forms that suit it.
Questions
- What has the collective actually shipped?
- A blockchain parametric flight-delay insurance product built at AXA and taken to market in two and a half months, on a group digital and big-data estate that was industrialised into a software factory in the same mandate. A conversational AI platform at Orange Business carrying 149 user paths. An AI conversational recruitment SaaS built from scratch at Peetchr, carrying a French DeepTech label.
- What does the AI work depend on?
- The data werk, always. A model on an ungoverned estate produces confident output from unverified input, which is worse than no model at all — it launders a data problem into a decision. AI mandates therefore start on the data platform and the lineage, not on the prompt.
- What is deliberately not claimed here?
- MLOps tooling and fine-tuning are declared, not proven: they are in scope for a mandate and no published case carries them yet. Model research is out of scope entirely. The collective applies and governs models; it does not build them.
- What does this werk actually build?
- It builds the machinery around the model rather than the model itself: the tool contracts an agent is allowed to call, the gateway every call is forced through, the guardrails that run before the prompt leaves the building, the evaluation harness that decides whether the thing is fit to ship, and the audit trail that lets a regulator reconstruct a decision months later. The model is the smallest and most replaceable part of the system.
- Why does the gateway matter more than the model?
- Because governance is only enforceable at a chokepoint. Prompt-injection detection, PII masking before inference, cost routing and lineage logging are each trivial to implement once and impossible to enforce across forty scattered call sites. On the On-Kare platform every model call routes through a single governed provider, and a CI gate fails the build on any call made outside it — with no baseline and no exemption list, because the first exception normalises the second.
- How is autonomy bounded?
- By tiering it against blast radius before a line is written. A tier-zero route — one that can end a treatment, block a prescription or trigger an emergency — gets hard stops, dual control, a hold-to-confirm and no auto-dismiss. A tier-three route gets ordinary product ergonomics. The tier is declared in the file name, so a reviewer cannot miss which class of hardening applies.
- Where is the proof today?
- Two client mandates and one venture. Orange Business shipped 149 user paths including 79 dialog flows on a conversational platform, with the team scaling from 20 to 30 during delivery. Peetchr was built from scratch as an AI conversational-recruitment SaaS running on social messengers, carrying a French DeepTech label. The deeper agentic engineering — gateway, guardrails, evaluation harness, lineage — currently lives inside the On-Kare venture rather than in a client case, which is why most of this werk reads HELD rather than PROVEN.