Work · Engagements in production

The systems we built. The accountability we signed.

3 engagements are published below in 6 parts, from challenge to evidence, with what was measured, anonymized at each client’s request. Every engagement carries a named partner.

Book a 30‑min call

Engagements in production

3 engagements. Anonymized. Walk‑through under NDA.

A named partner facilitates the briefing, presents the architecture and answers the questions a board, a regulator or an auditor would ask. Scheduled quickly with the named partner.

Luxury · Global maison

Cross‑border digital sovereignty for a global luxury leader.

Challenge: Keep the brand’s digital presence in mainland China under local data residency, without breaking brand consistency across continents.
Ambition: One brand experience across 2 continents, under one governance.
Transformation: Cross‑border data residency and cross‑border brand governance, sovereign by construction, with an audit trail by architecture.
Execution: Data residency engineered in for mainland China, one governance across both jurisdictions, and a governed transfer between them, with its record.
Outcome: The brand’s digital presence secured across mainland China, with brand consistency and operational reliability preserved across 2 continents.
Evidence: Walk‑through with the named partner, under NDA, on request. No performance figure is published.
What was measured: Indicator: residency of client data in mainland China, and the operational reliability of the brand’s digital presence across 2 continents. Measured by the client; values not disclosed. Method: record-based, from the record of each governed transfer between the 2 jurisdictions. The value is not published because the client asked for anonymity.

Cross‑border data residency
Sovereign by construction
Audit‑trail by architecture
Cross‑border brand governance

ASystems resident in mainland China, beside the boutique
BThe brand's systems in its other jurisdictions, beside the atelier
CThe governed transfer between them, with its record

Text version

2 client plates. On the first, mainland China: a boutique and the systems that keep data resident, inside a dashed residency frame. On the second, the brand's atelier and its systems in other jurisdictions. One dashed governed transfer links the two, and its record sits on the brand's plate.

Financial services · European tier-1

Production AI workflows under model risk discipline.

Challenge: AI‑augmented decision support in a regulated customer workflow, with DORA in scope.
Ambition: Run that workflow in production under model risk discipline.
Transformation: Model risk governance integrated from the architecture phase, with multi‑model orchestration.
Execution: Audit trail emitted at runtime; a model swap is a routing change, with the audit unchanged.
Outcome: Production AI workflows whose evidence is emitted, not assembled.
Evidence: Walk‑through under mutual NDA. No performance figure is published.
What was measured: Indicator: completeness of the audit trail emitted at runtime, and a model swap made as a routing change with the audit unchanged. Measured by the client; values not disclosed. Method: record-based, on the audit trail itself. The value is not published because the engagement is under mutual NDA.

Audit by architecture
DORA in scope
Model risk discipline
Multi-model orchestration

AThe bank's core ledger and the decision support agent
BAudit records, emitted at runtime, the latest highlighted
CThe regulatory obligations the records feed

Text version

A client plate holds the bank's core ledger, a decision support agent and a chain of audit records, the latest highlighted. A thin line runs from the chain to 2 regulation documents beside the plate: the evidence is emitted while the workflow runs, not assembled later.

Healthcare · European medtech

Production AI with human‑in‑the‑loop checkpoints.

Challenge: An AI workflow under European medical device constraint, with MDR in scope.
Ambition: Production AI with human‑in‑the‑loop checkpoints.
Transformation: Human oversight integrated into the architecture, not bolted on, with a kill switch by architecture.
Execution: Oversight designed into the workflow; audit pack emitted at runtime.
Outcome: The AI workflow shipped under European medical device constraint.
Evidence: Walk‑through under mutual NDA. No performance figure is published.
What was measured: Indicator: human oversight checkpoints applied in the workflow, and the audit pack emitted at runtime. Measured by the client; values not disclosed. Method: record-based, from the audit pack emitted at runtime. The value is not published because the engagement is under mutual NDA.

MDR in scope
Human‑in‑the‑loop
Audit pack at runtime
Kill switch by architecture

AThe clinician who oversees the workflow
BThe approval checkpoint on the oversight line
CThe patient record and the medical device
DThe agent that runs the workflow
EThe audit pack, emitted at runtime

Text version

A hospital plate holds the patient record, a medical device, the agent that runs the workflow and a chain of audit records. A clinician stands above the agent; an oversight line runs from the clinician down to the agent through an approval checkpoint.

Proof doctrine. Cases are anonymized at each client’s request, and no performance figure is published. Named clients appear as logos only, with their consent. The walk-through of a case takes place under NDA, led by a named partner. No testimonial is published without the author’s name, role and consent. The testimonial on the About page is published with its author’s consent, recorded on 25 October 2026. 2 named precedents stay in text because a public record supports each: the 2018 Crédit Mutuel cohort, published with its video and transcript, and the Numendo introduction, by a client named among the logos. Every operational claim points to the artefact that makes it verifiable: the 9-field record, controls C1 to C4, the 6-test due diligence and the 15 GOV controls.

Read the 9-field record and controls C1 to C4

Named clients · software and platform engagements

Clients who agreed to be named, on software and platform work. Our regulated‑sector engagements run anonymized, above; references on request, under NDA and with each client’s consent.

What we sign for

4 properties define every engagement we sign.

They are contractual, not aspirational. They are why the trust pack exists and why the comparative page is published.

01 · Named partner accountabilityOne of 5 partners signs the engagement and remains accountable through delivery, run and exit. No delegation. No junior substitution.
02 · Ownership transfer from day oneArchitecture decisions, runbooks, kill switches and audit trails belong to the client at exit. No black box. No key‑person dependency.
03 · Run-grade from line oneSLOs, audit trail, kill switch and human‑in‑the‑loop checkpoints are part of the build, not added before the audit.
04 · Audit-defensible by architectureEvidence is emitted by the platform itself. EU AI Act evidence by architecture, ahead of the December 2027 high‑risk deadline. DORA, NIS2, HIPAA, MDR scoped per engagement.

Ownership at exit

What you keep at exit.

Custom code, custom models, fine-tunes, prompt libraries, runbooks and operating documentation belong to the client at exit. Reusable components we retain are listed exhaustively in the MSA, so nothing arrives as a surprise licence.

ACustom code, in your repository
BCustom models and fine-tunes
CPrompt libraries
DRunbooks and operating documentation
EYour team, which operates the layer after transfer
FReusable components we retain, listed exhaustively in the MSA
GThe named partner who signs the engagement through exit
HThe transfer of ownership at exit

Text version

2 plates. On the client's plane: the custom code repository, the custom models, the prompt libraries, the runbooks and 2 client operators who run the layer after transfer. On the Hikari Blue plane: the reusable components we retain, the MSA that lists them and the named partner. An arrow marks the transfer of ownership at exit.

Every element shown is part of the engagement; none is decorative.

Next step: test these claims against a model we do not control. Verify Hikari Blue

Bring your case

A named partner reads your brief.

Describe the workflow and its principal constraint. A named senior partner replies quickly, in writing. If the brief fits, the next step is an Operating Assessment: 1 week, from $18,000.

Book a 30‑min call

Ask Hikari helps you frame the problem and proposes a first move. From the overlay, you can send the brief to a named partner with your name and work email.