This page (nudge.sterngold.nl) serves a public cloud demo of Nudge.
It runs in a Cloudflare Worker and calls a hosted language model directly. It is intentionally
light, fast, and shareable. It is not the system Milena and Vlad use with paying clients.
The production consultant product runs locally on a single Mac (AndersStar) over a private network (Tailscale). A practitioner's voice profile, methodology corpus, and client conversations never leave that machine. The cloud demo here cannot make that claim — and that is the point of the distinction.
| Property | This cloud demo | Consultant product (LOCAL) |
|---|---|---|
| Where the model runs | Cloudflare Worker → hosted LLM (Anthropic) | Local machine (Ollama + a fine-tuned voice model on AndersStar) |
| Who can reach it | Anyone on the open internet | Only the practitioner, over a private network (Tailscale-only) |
| What's logged | Session content kept up to 90 days in Cloudflare KV (so we can review demo behavior) | A per-session audit log on the practitioner's machine; no third-party telemetry, no analytics |
| Voice fidelity | Generic — no practitioner-specific tuning | Per-practitioner LoRA adapter trained on their own materials |
| Safety gating | Light — surfaces the same look-and-feel as production | Three-zone Safety Monitor (Green ships / Yellow ships with team recall / Red aborts) |
| Data boundary | None claimed | Vault chunks, methodology, and client conversations never leave the local machine |
Use the cloud demo to feel the shape of the product. If you want to see what your practitioner would actually deliver to your operation, ask them for a session run on the consultant product — that's where the voice, methodology, and safety contract are real.
Use the consultant product (the LOCAL — AndersStar bundle, served on your Tailnet from your own machine). Don't use this cloud demo for paying clients. The badge on every page of the demo — Cloud demo — not for client work — exists to keep that line bright.
The architectural separation between the two paths is documented in the Nudge engine spec: two physically separate frontend bundles, one served by a public Worker, the other served only by a local FastAPI process over a private network. Source repository: sterngold/nudge.