AISE · Xerj demo · v4

~3 min 45 sec · 1440×900 · captured 2026-04-28 · every CLI line, every dashboard number, every passkey ceremony — real-time on this machine

↓ Download MP4 (9 MB) Press space to play/pause; ←/→ to seek 5s; click the timeline to scrub.

Six scenes, ~225 s total

Time	Scene	What's on screen
0–8s	Setup tooltip	Xerj is the search engine. Xerj Console is the typography-first UX bundled inside the same binary. One file, no JVM, no separate Kibana to install.
8–55s	Launch Xerj	`ls -lh xerj` — 22 MB binary. `./xerj --insecure --data-dir …`. Boot log highlights `xerj v1.0.0-rc.1 starting`. Then the boxed XERJ CONSOLE · first-launch setup banner — tooltip explicitly notes this is Xerj Console's claim flow, the server is already up. Proof: `curl /` returns `cluster_name=xerj`, `/_cluster/health` = green, `ss -lntp` shows ports 8080 (native) + 9200 (ES-compat) listening.
55–80s	Show data file	`head -1 chat-events.ndjson \| jq` reveals the LLM-telemetry shape — `model: claude-haiku-4-5`, `prompt_tokens`, `cost_usd`. Same model name you'll see on the dashboard's BY MODEL chart at scene 6.
80–115s	Bulk ingest	Single visible `curl POST /_bulk` for chat-events. Response: `took:104ms · errors:false · n:4008 · first:201 · last:201`. Verified with `_cat/indices/chat-events` showing 4 008 docs. Off-screen the orchestrator pushes the other five corpora silently so dashboards have data.
115–145s	Verify	Run the EXACT aggregation the AI Overview adapter sends: `queries:4008 · tokens_M:44 · cost_usd:38 · models:{sonnet:1390, haiku:1110, opus:889, gpt-5:372, gemini-3:158, llama-4:90}`. Bucket sum = 4 008 = full corpus.
145–165s	Open Xerj Console	Browser opens magic-link, types `[email protected]` + `Demo Owner`, virtual WebAuthn enrols passkey, redirect to `/_xerj-console/`.
165–225s	Walk Xerj Console UX	DATA — six ingested indices visible on LOCAL XERJ cluster, 12.4 M docs total DISCOVER — interactive query console AI · OVERVIEW — 3.7K queries, 40.4M tokens, $35 (numbers slightly different from verify scene because of 24h time-filter) RAG · QUALITY — same data, different shape: hybrid recall, citation rate, top docs VECTOR · INDEX — 6.39M ops, 1536-d HNSW, 0.94 recall AGENT · MEMORY — 1.78M entries, 32.3% dedup rate, 78.3% recall, 8 agents ALERTS — rule-based + anomaly destinations SETTINGS — Demo Owner / OWNER role / passkey only / dashboard list

Three points the v2 demo answers explicitly

Why does the boot banner say "XERJ CONSOLE"? Because Xerj Console (the bundled UI) prints its own first-launch claim banner. The server (Xerj) prints its own version line one paragraph above; v2 highlights it.
Where do the dashboards come from? Eight templates ship as code under xerj-ux/src/dashboards/*.js, bundled into the binary at compile time. The data they render comes from whatever you POST to /_bulk; users can rename / hide / clone via the SETTINGS section but the templates are versioned with the code.
How is the data added to a dashboard? It isn't — pull-based. Drop docs into the chat-events / vector-ops / agent-memory / anomalies / ingest-events / ai-kb indices and the matching dashboards' adapters auto-pick them up via _search aggregations. No "connect data source to dashboard" step required.