Willkommen bei PULSE Fitness.
Sprich mit Alex, deinem AI-Coach am Empfang — über Kurse, Geräte, Tarife und dein kostenloses Probetraining. Auf Deutsch oder Englisch, wann immer du willst.
Sprich mit Alex, deinem AI-Coach.
Dein Mikrofon wird erst beim Start des Gesprächs angefragt. Das Video kommt direkt vom Avatar-Anbieter zu dir — wir sitzen nicht im Latenzpfad.
Live transcript
Three pivots in three days.
Every pivot taught us something. None of the work was thrown away — the engine, the dashboard and the deploy stack survived all three.
-
Watchstore prototype — a sandbox to wire Anam
Started with an imaginary luxury watch boutique called "Daniel's Watch Store" to learn the Anam SDK end-to-end. First-light: WebRTC stream, mic capture, English-only persona "Marta". Caddy + Hetzner + systemd plumbing came together here.
-
Pivot to Araibia / Zaid — own marketing demo
Reused the entire stack but swapped the persona, knowledge and UI to introduce Araibia itself. Three knowledge files (about, use cases, Vision 2030), light lavender palette, a single agent showcase card. The live URL stayed the same.
-
Pivot to Rolex Boutique — first real client brief
Read the full Rolex Avatar PRD. Locked the architecture: stay on Anam End-to-End for now, queue Brain-Mode for Phase 2. Replaced the knowledge base with 15 Rolex catalogue files plus brand essentials, service, boutique info, and conversation patterns. New persona "Khalid" — Saudi male, 30, polished. PRD §6 system prompt baked in, with all seven hard rules.
-
This page — the build log itself
The status report became the public site. Live capabilities, pending hand-offs, future vision, and a comment section so the people who care about this can leave feedback in real time.
Live and verified.
Every item below is on the production server, restartable as a systemd service, and proven by a real request in this build log.
Public site behind HTTPS
avatar.danielstradingbot.xyz, served by Caddy with auto Let's Encrypt. Hetzner CX22 in Frankfurt, Ubuntu 24.04, systemd-managed Hono service on Node 22.
Anam-WebRTC avatar streaming
Session-token endpoint authenticates server-side; the video stream goes from Anam's cloud directly to the visitor's browser. Our backend doesn't touch audio or video frames.
Knowledge base — 321 chunks
15 Rolex catalogue families plus brand essentials, service rules, boutique info and conversation patterns. Stored as markdown in the repo, indexed at boot, retrievable in well under 5ms with BM25.
Admin dashboard
Basic-auth-protected. Persona-Helper auto-generates the PRD-conformant system prompt; Knowledge-Export bundles all chunks for paste into Anam; RAG-Test verifies retrieval; Session-Log shows every token request.
Persona prompt with seven hard rules
No invented facts, no prices quoted, no stock promises, no investment advice, no competitor disparagement, no PII collection, no off-topic drift. All compiled into the live system prompt from PRD §6.
This page · with comments
The build log is itself a deployed feature. Comments persist to disk on the server, rate-limited per IP, XSS-safe on render.
Two final hand-offs before Khalid greets a customer.
The plumbing is done. What's left is configuration on Anam's side and a UI pivot on ours.
Anam dashboard configuration
Open /admin, copy the auto-generated system prompt and the knowledge export, paste both into the Rolex persona on Anam. Pick a bilingual voice. Add the domain to the allow-list. Click Publish.
Public-facing pivot to Rolex Boutique
The hero, brand mark and copy still read "Araibia / build log". For the boutique deployment we'd swap to the Rolex narrative on a separate subdomain, keep this one as the project page.
From third-party stack to sovereign agent platform.
Where this is going. Six phases, each builds on the last, none of them speculative — every layer is something a serious agent business needs to own.
-
02Quarter 2
Brain-Mode — own the LLM
Replace Anam's default LLM with a direct call to Claude Sonnet 4.5 (or equivalent). Stream tokens through our backend so we own the conversation log, the guardrail enforcement, and the hand-off logic. Anam becomes the mouth, face and ears; we are the brain.
-
03Quarter 2
Real RAG — vector database with multilingual embeddings
BM25 is unilingual. Replace it with pgvector or a managed vector store on top of multilingual embeddings (text-embedding-3-large or Cohere multilingual). An Arabic question now retrieves from English source content correctly — non-negotiable for Saudi customers.
-
04Quarter 3
Multi-tenant agent platform
One platform, many agents. Each client gets their own persona, their own knowledge silo, their own analytics — all running on shared Araibia infrastructure. Agent-to-knowledge linking means a single knowledge graph can power any number of agents, with permissioned views.
-
05Quarter 3
In-region infrastructure — own servers in Saudi
Migrate from generic European hosting to AWS me-south-1 (Bahrain) or directly to Saudi-resident infrastructure. Latency drops, data residency clears, and the agent feels distinctly local — both technically and culturally.
-
06Quarter 4
Central knowledge backbone — agents share a graph
A live knowledge graph linking products, brands, customers, intents and conversation patterns. Agents traverse the graph in real time. New product launches propagate to every agent automatically. Every conversation feeds back into the graph as structured signal.
-
07Year 2
Sovereign avatar rendering
The bold one. Move avatar rendering and lip-sync in-house, using the latest open generative-video models. No vendor lock-in to Anam or HeyGen, full control over Arabic phoneme handling, full ownership of the customer-facing face. The end state of "Araibia owns its stack."
What would you build next?
Thoughts on what's working, what's missing, what should ship next quarter — leave them here. Public, persistent, no account needed.