CAPABILITY · 01 · AI INTEGRATION
We build AI into the systems your licensed people already use — not on top, not instead. Every recommendation cites the data it leaned on. Every model has a card. Every output can be rolled back.
ARCHITECTURE
The same AI fabric runs across all four verticals. Vertical signals go in, inspectable decisions come out, the licensed human keeps the seat of judgement.
INSTITUTIONAL
anomaly · case triage · audit narrative
Case #1284 — sources cited (3)
SUPPLY CHAIN
ETA · variance · counterparty signal
ETA 0.91 — telemetry cited (4)
HEALTHCARE
ambient docs · CDS · triage tier
Note drafted — guideline cited (2)
FINTECH
fraud · AML triage · treasury
SAR draft — rules cited (5)
WHAT WE BUILD
01 · VALIDATED RISK
Independent statistical review before production. Performance, fairness, drift — monitored continuously and stamped on the audit page. If a model drifts past your tolerance, it pauses itself and pages your team.
02 · AMBIENT EVIDENCE
Ambient capture for consults, ops calls, treasury reviews. The model drafts the note, the licensed human signs it. The recording goes to your tenant, never to a third-party model trainer.
03 · CITED DECISIONS
Hover a decision, see the documents it leaned on. Sources are not a debug button — they're a regulatory artefact. If we can't cite the source, we don't ship the recommendation.
AI · BY VERTICAL
We talk about AI the way operators do — in terms of the move it makes. Here's what it looks like in each vertical we ship.
INSIDE A RECOMMENDATION
This is a representative slice of the case detail your team would see. Click a citation marker to expand the source it pulled from. Click again to collapse. No 'why did it say that' guessing.
The applicant matches the profile of approved customers in adjacent risk tier with the following deltas:
FILING REGISTRY · 2024-2026 RECORD
Annual filings present and timely. No restatements. Auditor unchanged.
FINAL DECISION HELD BY · COMPLIANCE OFFICER ON DUTY
WHAT WE'VE WRITTEN DOWN
We publish what we've learned. Three short notes that institutional buyers ask us for in the first call.
WHITE PAPER · 01
The one-page model card we ship with every production model. Inputs, performance, fairness, drift, owner, expiry. The artefact your regulator opens first.
Read note
WHITE PAPER · 02
Why we make every AI model pass an external statistical review before production — and what that review actually covers. The template is in the appendix.
Read note
WHITE PAPER · 03
How we build the AI parts of the platform so they're modular, replaceable, and don't lock you to our infrastructure. Sovereignty applies to AI too.
Read noteWHITE PAPER · 01
The one-page model card we ship with every production model. Inputs, performance, fairness, drift, owner, expiry. The artefact your regulator opens first.
Read noteWHITE PAPER · 02
Why we make every AI model pass an external statistical review before production — and what that review actually covers. The template is in the appendix.
Read noteWHITE PAPER · 03
How we build the AI parts of the platform so they're modular, replaceable, and don't lock you to our infrastructure. Sovereignty applies to AI too.
Read noteAI · COVENANTS
Other vendors put these in the appendix. We put them in the contract. Each one is enforceable, audit-checkable, and yours to invoke at any time.
CITATION · ALWAYS
If a model output reaches a human screen, the documents and rules behind it are one click away. No 'black box' UI states. No 'come back later for the explanation.' If the trace isn't there, the recommendation doesn't render.
REVERSIBLE · BY DEFAULT
AI doesn't write to your system of record directly. It proposes — your licensed human accepts. The accept itself is logged, the rollback is one click, and the data behind both is preserved.
MODEL · CARDED
Inputs, training-set provenance, performance band, fairness band, drift policy, model owner, expiry date. Card is a contractual artefact — if it's missing, the model isn't licensed to run.
NEVER · AUTONOMOUS · MONEY
AI can recommend the wire, populate the SAR, draft the consent form. The human signs. The two-person rule never goes away because of a model's confidence score.
DATA · COVENANT
We train on your data only with your written, time-bounded consent. We never use your data to train a model another customer benefits from. Inference happens inside your tenant. We don't see your prompts. We don't see your completions.
TENANT-LOCAL
Prompts, completions, traces — all stay in your environment. Hylen has no read-access to runtime AI traffic, by design and by contract.
CONSENT-BOUNDED
If we use your data to train a model, the consent is written, time-limited, and per-purpose. No 'general improvement' clauses. No silent retraining.
NO-CROSS-TRAINING
Models trained on your data serve you only. We do not pool, anonymise-and-mix, or 'pre-train and fine-tune' across customer tenants.
VENDOR-MODEL · DISCLOSED
If a recommendation leans on a vendor model, the model name, version, and the data it sees is in the audit log. No silent dependencies. No vendor lock-in by surprise.
WHY THIS MATTERS
“AI does its best work when the human doing the actual job can see what it's looking at, what it's leaning on, and what it just changed. Anything else is a model talking to itself.”
Operating principle · 04
STATUS · OPEN
We have capacity for two new AI integration engagements before quarter-close. First conversation is on us. We respond within four business hours.
