Appearance
AI Protection Admin — Sales Engineer Demo Script
Runtime: about 14 minutes · Format: six live demos in the org admin AI Protection area (API or admin UI, depending on your build)
This script positions GhostAI AI Protection Admin as the control plane for where inference runs, how PII is verified, how the AI referee is configured, and how embeddings can be rebuilt—without asking buyers to read Python or infrastructure runbooks.
Pre-demo checklist
Complete before you share your screen:
- [ ] Org admin or owner account signed in (AI Protection routes require
require_admin; members and viewers receive 403). - [ ] API key or session ready if you are demoing via
curl/REST client alongside the UI. - [ ] Private model stack (Ollama or vLLM) optionally running so
private_model_availableandinference_targettell a positive story; if it is down, use the script’s “honest failure” lines in Demo 1. - [ ] Bedrock credentials and region configured on the server if you plan to show referee model discovery or Bedrock-backed referee health (otherwise Demo 4–5 gracefully show empty models or local-only runtime).
- [ ] Sample org traffic (or staging data) so metrics for 7d/30d are non-zero if the buyer expects charts.
Opening line (15 seconds): “Every message hits PII detection and optional AI verification before it reaches a model. AI Protection Admin is where security leaders see that posture and set org-wide defaults users cannot override when you lock them.”
Demo 1 — Protection Status Dashboard (~2 min)
Navigate: Admin → AI Protection → Status (or GET /api/admin/ai-protection/status).
Say: “This is the single pane of glass for private inference readiness and NER model status for this tenant.”
Walk through the payload:
| Area | Fields to narrate |
|---|---|
| Private model | private_model_available — boolean rollup of load/probe success. |
| Private model detail | private_model_status: model, runtime, deploy_label, loaded. |
| Inference target | inference_target: location, url, model, models_available, gpu, gpu_memory_gb, error, tls, tls_warning. |
| Cloud default (read-only) | default_cloud_model — server default when policy allows cloud. |
| NER model | Dymium NER (dymium/Dymium-NER-v1) — the DeBERTa-v3-large NER model used for entity detection. |
If error or tls_warning appears: “The platform is transparent: you see TLS and endpoint errors here instead of silent fallback in the UI.”
Bridge: “Next we bind this infrastructure to policy—who must use private inference and whether users can turn verification off.”
Demo 2 — Org Protection Policy (~3 min)
Navigate: Admin → AI Protection → Policy (or GET/PUT /api/admin/ai-protection/policy).
Say: “Org policy sets defaults and locks. Locks are how regulated teams stop individual users from opting into a weaker posture.”
Explain each control:
| Setting | Behavior |
|---|---|
default_ai_model | cloud → stored as user (end-user cloud selection). private → local. custom → dedicated. |
lock_ai_model | When true, users cannot override the org’s default AI model choice. |
default_verify_mode | all — verify broadly; suspicious — targeted verification; off — no AI referee verification (still subject to product rules). |
lock_verify_mode | When true, users cannot change verify mode away from the org default. |
pii_gliner_model_path | Deprecated. Formerly the per-org custom GLiNER checkpoint path. NER now uses dymium/Dymium-NER-v1. |
Live action: Toggle default_ai_model to private, enable lock_ai_model, set default_verify_mode to all, enable lock_verify_mode. Save.
Audit hook: “Every material policy change is written to the admin audit trail as ai_protection_policy_updated with the fields that changed—we will show that in the wrap-up if the buyer cares about SOC2 evidence.”
Demo 3 — PII Detection Metrics (~2 min)
Navigate: Metrics (or GET /api/admin/ai-protection/metrics?period=30d).
Say: “This is org-scoped telemetry derived from protection activity—not marketing counters.”
Cycle the period query: 7d, 30d, 90d.
Highlight:
total_sessions_protected— distinct sessions with at least onetokenizedaudit event in range.total_entities_detected— volume of tokenization events (entities detected in the audit sense used here).entities_by_type— array of{ entity_type, count }for executive-friendly breakdowns.- Model split —
detection_model_usagemaps friendly names:private(local),cloud,custom; plususers_with_private_model,users_with_cloud,users_with_custom. user_overview— per member:ai_model(friendly),matches_org_default,entities_detectedin the period.
Close: “This answers ‘are we actually on private?’ and ‘what did we catch?’ in one response.”
Demo 4 — Referee Configuration (~3 min)
Navigate: Referee configuration (or GET/PUT /api/admin/ai-protection/referee/config).
Say: “The referee is the LLM that resolves ambiguous PII calls. Admins can steer runtime and Bedrock without redeploying the whole app—hot-swap is built in.”
Discovery first: GET /api/admin/ai-protection/referee/models — Bedrock model IDs available for the configured region (uses server-side discovery). If region is unset, the response explains that with message or error.
Current config: GET /api/admin/ai-protection/referee/config returns:
runtime— e.g.ollama,vllm,bedrock.bedrock_region,bedrock_model,bedrock_endpoint_url.fallback— referee fallback mode.audit_bedrock— whether Bedrock referee calls are auditable per server settings.max_concurrent— concurrency cap for referee work.
Live action: PUT a minimal change the environment allows (for example max_concurrent or bedrock_model in a sandbox). Say: “The server calls rebuild_referee after updates; the JSON response includes an updated object and, when the swap runs, a status field reflecting readiness or failure.”
Caution line for enterprise buyers: “Referee settings are process-wide; in production you coordinate changes with the platform team so you do not surprise other tenants if you share a control plane.”
Demo 5 — Referee Health (~2 min)
Navigate: Referee health (or GET /api/admin/ai-protection/referee/health).
Say: “Status complements the dashboard: this is referee LLM load state and resilience signals, not duplicate GPU telemetry.”
Point to:
runtime,status,loaded— derived fromrequest.app.state.pii_referee_llmandpii_referee_llm_status.fallback— current fallback setting.- When
runtimeisbedrock, thebedrockobject:region,model,audit_enabled. metrics.circuit_state— referee circuit breaker gauge when metrics are available (non-zero values indicate tripped or degraded states depending on your Prometheus mapping).
Differentiator: “For inference URL, GPU, TLS, buyers still use Demo 1 (/status). This endpoint is referee-specific.”
Demo 6 — Embedding Rebuild (~2 min)
Navigate: Embeddings / maintenance (or POST /api/admin/ai-protection/embeddings/rebuild).
Say: “When you change embedding backends—local sentence-transformers versus Bedrock—you can re-vectorize stored memory and content-library text in place.”
Request body (JSON):
target_backend:localorbedrock.scope:all(memory + content-library),memory, orcontent-library.batch_size: integer concurrency hint (default 100 in API).
Live action: Start a job; capture job_id from the response (status: started).
Poll: GET /api/admin/ai-protection/embeddings/rebuild/{job_id} for status (running, completed, or failed: …), total, completed, failed, target_backend, scope.
Set expectation: “Large corpora take time; this is intentionally asynchronous.”
Close (~1 min)
Recap: Status (private + NER + inference target) → Policy (defaults + locks + audit ai_protection_policy_updated) → Metrics (period + entities + model split) → Referee (discovery, config, hot-swap) → Health (load + Bedrock + circuit) → Embedding rebuild (job tracking).
Ask: “Do you want governance in the admin audit log next, or a deeper architecture session on tokenization versus referee?”
Appendix — Quick reference (API)
| Purpose | Method | Path |
|---|---|---|
| Status | GET | /api/admin/ai-protection/status |
| Policy | GET, PUT | /api/admin/ai-protection/policy |
| Metrics | GET | /api/admin/ai-protection/metrics?period=30d (supported: 7d, 30d, 90d) |
| Referee models | GET | /api/admin/ai-protection/referee/models |
| Referee config | GET, PUT | /api/admin/ai-protection/referee/config |
| Referee health | GET | /api/admin/ai-protection/referee/health |
| Start embedding rebuild | POST | /api/admin/ai-protection/embeddings/rebuild |
| Embedding job status | GET | /api/admin/ai-protection/embeddings/rebuild/{job_id} |
All routes require org admin authentication (same bearer token pattern as other /api/admin/* endpoints).