Observer/Dashboard
⌘Kquick switch[email protected]
OVERVIEW · 00

Compliance help layer
?
What is Observer?A non-invasive help layer for compliance apps. The Observer reads what users see, the Knowledge Engine maps it to regulatory concepts, the AI Generator writes plain-language guidance, and the UI Injector shows it. Host apps stay untouched.

Watches your host apps, maps form fields to regulatory concepts, generates compliance-grade help, and injects it into the UI — without touching the host application's code.

KB VERSION
v0.1.0-samples
5/12/2026

The pipeline

Observer → Knowledge → AI Generator → UI Injector
Observer
reads route + identity + DOM
Knowledge Engine
field → concept
AI Generator
Claude Sonnet 4 + grounding
UI Injector
shadow-DOM render
Host UI
user sees help
Host app never imports a single line of help-layer code. Pure injection.

Service status

live probes
registry-api
45ms
knowledge-api
45ms
content-api
53ms
LLM mode
live
Portfolio
?
PortfolioEvery host app the Observer can be deployed to. observerEnabled: false keeps the bundle loaded but silent — a safety net for rollout.
37
0 observer-enabled · 37 silent
Knowledge base
?
Knowledge baseConcept entries authored by your domain experts. Only published concepts surface in live resolution; drafts are loadable but invisible.
3
3 published · 0 draft
Telemetry (24h)
?
TelemetryHelp:* events streamed in from every Observer instance. Sessions = unique HelpContext senders. Errors = help:error events captured.
0
0 events · 0 errors
Content gen (24h)
?
Content generationClaude Sonnet 4 turns each KnowledgePacket into a ContentPayload. Prompt caching keeps cost down; degraded mode keeps the lights on without an API key.
1
687ms avg · 0 cached
Rollout
?
RolloutStage gates for production deployment across the host-app portfolio. Each stage represents one app going live.
1/6
17% complete
Coverage
?
CoverageHow much of your compliance surface area is documented. One row per knowledge domain (AML/KYC, tax, AI governance, etc.).
1
3/3 concepts ready
Cache
?
CacheActive ContentPayload cache entries. Keyed on appId+conceptId+role+jurisdiction+language+flags+KB version. 24h TTL standard, 1h for degraded entries.
0
24h TTL · SHA-256 keyed
Live mode
?
Live modelive = Claude Sonnet 4 is reachable. degraded-no-llm = ANTHROPIC_API_KEY missing or Anthropic API down. Either way, users still get accurate help — only the polish differs.
✓
live

Get started
?
What's the order?Each step is independent — you can skip ahead. But this is the order that gives you a live demo fastest. The whole sequence takes ~30 minutes once you have an API key and a willing host app.

2 / 5 done
✓
Register your host apps
37 registered
→
✓
Author at least one concept
3 published, 0 draft
→
3
Enable Observer on a staging app
0 enabled (all silent)
→
4
Try the live Inspector end-to-end
build a synthetic context → Resolve + Generate
→
5
Deploy Observer + Injector to a host app
two <script> tags, zero host-app changes
→

What lives where

Observer + UI Injector are vanilla-JS bundles you drop into your host apps as <script> tags. Zero deps, <8 KB gzipped.

Registry, Knowledge, Content APIs are this server. They never touch the host app directly — the bundles call them via HTTP.

This console is for you, not your end-users. It manages the portfolio, authors the knowledge base, and observes telemetry.

Why does this exist?

Compliance apps have the worst documentation problem in software. AML/KYC fields exist because of specific regulations, with specific consequences for getting them wrong, and the rules differ per jurisdiction and per role.

Generic help tools answer "what does this button do." This one answers "why does this field exist, what happens if I get it wrong, and what should I do next."

See the help section for the full architecture.