Skip to main content

Architecture Overview

Adjudon is a Decision Audit Layer for regulated AI deployments. Three concerns sit on the customer side of an HTTPS call — agents, decisions, identifiers — and three sit on ours: a confidence engine, a policy engine, and a tamper-evident SHA-256 hash chain. The seam between them is the only contract that matters.

The shape of the system​

Three production surfaces:

SurfaceURLStack
Landing siteadjudon.comNext.js 14, Cloudflare Pages
Dashboardapp.adjudon.comReact 18 + Vite, Cloudflare Pages
APIapi.adjudon.comExpress 5, Fly.io Frankfurt

The API hosts every customer-facing endpoint — about 42 route files, roughly 260 endpoints. The dashboard reads from those same endpoints; the landing site has no access to customer data. Every customer integration crosses one HTTPS call: POST /api/v1/traces records an AI decision, the chain hashes it, and the response returns the row's chainHash. Reviews, alerts, webhooks, and exports are all downstream of that single ingestion.

The full request path:

                 ┌─────────────────────────────────────────────┐
                 │  api.adjudon.com  ·  Express 5  ·  Frankfurt │
                 │                                             │
 [Customer agent]│  PII scrub                                  │
   ──── POST ───▶│   ↓                                         │   async
 /api/v1/traces  │  Confidence Engine (3 pillars)              │ ────────▶  Webhook dispatch
                 │   ↓                                         │
                 │  Policy Engine ── 201 / 202 / 403           │
                 │   ↓                                         │
                 │  HashChainEntry append                      │
                 │   ↓                                         │
                 │  MongoDB Atlas Frankfurt ◀────┐             │
                 │                               │             │
                 │  GET /api/v1/hash-chain/      │             │
                 │   export   ───────────────────┘ self-       │
                 │                              contained      │
                 │                              JSON bundle    │
                 │                              (offline       │
                 │                               replay)       │
                 └─────────────────────────────────────────────┘

The pipeline holds tight latency SLOs: p50 < 10ms, p95 < 25ms, p99 < 45ms end-to-end through ingestion, scrubbing, confidence, policy, and chain append. The Express 5 server runs on Fly.io Frankfurt with health-check responses below 100ms independent of database load.

Where data lives​

ServiceRegionPurpose
MongoDB AtlasFrankfurt (eu-central-1)Primary database
Fly.ioFrankfurtAPI server
Cloudflare PagesEU edgeLanding + dashboard + docs
StripeIrelandBilling + metered usage
ResendEUTransactional email

Frankfurt eu-central-1 is the rule, end to end. The Confidence Engine's third pillar (memory similarity by vector embedding) runs on a self-hosted TEI (Text Embeddings Inference) sidecar in the same Frankfurt Fly.io region — the all-MiniLM-L6-v2 model is loaded into memory at startup and never makes outbound calls. OpenAI L.L.C. (USA) was removed from the sub-processor list on 2026-05-11; no customer trace payload leaves the EU for any reason. Cross-organization isolation is enforced in code: every DecisionTrace query filters by organizationId, no exceptions.

The Confidence Engine​

Confidence is triangulated from three independent signals:

  • Base probability, derived from the decision context, not from the model's own self-reported softmax score
  • Variance, computed across a small ensemble of model calls; high variance means the model disagrees with itself
  • Historical precedent, comparing the current decision against similar decisions in the organization's past trace history

The engine emits a single confidenceScore in 0.0–1.0 plus tags: LOW_CONFIDENCE when the score falls below 0.6, HIGH_AMBIGUITY when variance is high. Tags drive routing (review queue, policy gates, alerts); the score appears in the chain row. An MDCG 2019-11 notified body or a GDPR Art. 22 reviewer will not accept the model's own self-reported softmax as evidence; triangulation is the operational answer.

The Multi-Clock Incident Hub​

A single AI-related incident at a German bank rarely triggers one regulator. It can trigger up to five at once, each with its own clock:

RegulatorArticleEarlyIntermediateFinal
GDPRArt. 33——72 h
EU AI ActArt. 732 d10 d15 d
DORAArt. 194 h72 h30 d
NIS2Art. 2324 h72 h30 d
CRAArt. 1124 h72 h14 d

The Multi-Clock Hub runs all five as parallel countdowns off one detection event. Each IncidentClock carries regulator, articleRef, clockStartedAt, an array of checkpoints with evidenceTraceId, and a status of active, paused, completed, or breached. A breached clock is not deleted — it stays in the index with status: 'breached' so the post-incident audit can replay which deadlines were missed and when.

Replay without our login​

Every DecisionTrace is hashed into a per-organization SHA-256 chain row:

HashChainEntry
{
  "sequence":      17493,
  "prevHash":      "a3c1...0f9b",
  "payloadDigest": "9e2d...4c7a",
  "chainHash":     "f1b0...2dde",
  "createdAt":     "2026-05-05T10:14:22.317Z"
}

The chain formula is published:

chainHash     = sha256(prevHash || payloadDigest || sequence || createdAt)
payloadDigest = sha256(canonicalJson(traceView))

The chain is tamper-evident, not tamper-proof: a hostile insider who edits a stored entry breaks the next entry's prevHash link. Tampering is loud.

When audit is scheduled, the bank's compliance team requests the bundle:

curl https://api.adjudon.com/api/v1/hash-chain/export \
  -H "Authorization: Bearer $ADJUDON_API_KEY"

The response is a self-contained JSON document: every entry, every hash, every sequence number. The auditor recomputes each chainHash and compares to the stored value. Result: verified: true or brokenAt: <sequence>. No third state. The verification runs against the downloaded bundle. No Adjudon login, no Adjudon endpoint, no Adjudon network. The chain is valid evidence even if Adjudon disappears between the export and the audit.

What we do not do​

Adjudon is SaaS-only: no white-label, no customer-cloud, no on-premise. We do not sit in the request path between your application and your LLM — that category (Lakera, CalypsoAI, NeMo Guardrails) is a separate vendor type, and a regulated stack typically wants both layers.

We do not store raw trace payloads in logs (Cardinal Rule 4) or send customer data to external ML training (Cardinal Rule 6). We do not modify chain entries (Cardinal Rule 5) — GDPR right-to-erasure nullifies the source DecisionTrace payload fields, but the chain shell and its payloadDigest (computed before erasure) remain intact.

See also​