# Powered by RIG — agency-white-label A white-label AI front desk that **dental marketing agencies resell to their own book of practices**. Wholesale $250 per practice per month; the agency sets retail (typically $650) and keeps the spread. The buyer is the **agency owner**, not the dentist. Everything here is built for that conversation. --- ## What's in this directory | File | What it is | |---|---| | `app/` | The working agency console. Open `app/index.html` in a browser. | | `landing.html` | Conversion landing page aimed at agency owners. | | `icon.svg` / `icon-256.png` | Product mark. Hand-authored SVG, 512×512. | | `GTM.md` | ICP, triggers, 5-email sequence, 3 LinkedIn posts, objections, 10 prospect types. | | `PRICING.md` | Wholesale, retail bands, margin math, RIG-side unit economics, billing mechanics. | | `app-shot.png`, `app-shot-2.png` | Real screenshots of the console, used by the landing page. | ## Run it ```bash open ~/Developer/rig-offerings/agency-white-label/landing.html open ~/Developer/rig-offerings/agency-white-label/app/index.html ``` No build step, no dependencies, no server. Plain HTML/CSS/JS opened straight off disk. **Useful URLs:** ``` app/index.html#desk jump to the live front desk app/index.html#margin jump to the resale model app/index.html?say=do%20you%20take%20delta%20dental#desk open mid-conversation app/index.html?accent=%2333C48D&agency=Northline%20Growth preview a prospect's brand ``` That last one is a genuine sales tool: send an agency a link that opens the console already wearing their name and colour. --- ## What is real - **The AI is real.** The desk calls `http://100.67.126.117:8000/v1/chat/completions` — vLLM serving `Qwen/Qwen3-Coder-30B-A3B-Instruct` on the blackwell node (3× RTX PRO 6000) — directly from the browser. vLLM returns `access-control-allow-origin: *`, so this works from `file://` with no proxy and no key. Verified in a real Chrome session: sidebar reads *"fleet: blackwell live"* and replies come back in double-digit milliseconds. - **The system prompt is real.** Each reply is constrained by the selected practice's record: hours, accepted insurance plans, new-patient offer, PMS, and the escalation rules (clinical questions, anything over $1,000, and direct "are you a bot" questions all hand off to a human). - **Booking detection is real.** The model emits a `[BOOKED: Day Time]` token, the app parses it, writes an appointment, increments the practice's booked count, and logs an event. Seen working end to end. - **The white-label engine is real.** Changing agency name, colour, subdomain, patient-facing desk name, or voice re-themes the entire UI live via CSS custom properties, rewrites the patient widget preview, regenerates the embed snippet, and persists to `localStorage`. - **All the math is real.** Portfolio roll-up, per-practice spread, the churn-decayed year-one model, and the practice-side ROI are computed from the data, not hard-coded. Every figure printed in `GTM.md` and `PRICING.md` was recomputed against the same formula the app uses. - **Adding a practice is real.** It writes into state, updates the roster, the totals, and the picker. - **The offline fallback is real.** With no network the desk answers from a deterministic rules engine so a demo never dies in a prospect's conference room. The UI says plainly which mode it is in. ## What is stubbed Be straight about this on sales calls. - **No PMS write-back.** "Written to Dentrix" is a UI event, not an API call. Open Dental has a usable API and is the realistic first integration; Dentrix and Eaglesoft need a middleware bridge on the practice's machine. Nothing is connected today. - **No telephony.** Missed-call text-back is simulated in the browser. Real delivery needs a Twilio (or similar) number per practice, forwarding configured at the practice, and 10DLC registration. - **No backend, no accounts, no auth.** State lives in `localStorage`. Two people opening the app do not share data. There is no multi-tenant server, no database, and no billing integration. - **`embed.js` does not exist.** The Deploy tab generates a correct-looking snippet pointing at `https:///embed.js`. That file has not been built and the DNS target is illustrative. - **The sample book is invented.** Six practices with plausible names, cities, plans and volumes. No real patient data, no PHI, and nothing HIPAA-relevant is touched anywhere in this build. - **The proof story in `GTM.md` Email 3 is illustrative**, not a real customer case. There is no reference customer yet. Do not present it as one. - **Ollama nodes are unreachable from the browser.** rig-96gb, rig-36gb, rig-28gb and rig-256gb all answered fine server-side, but they return `403` on a CORS preflight from a browser origin, so they are used for batch generation only. blackwell is the only browser-callable node. ## Nearest real thing to build next 1. Twilio number + missed-call webhook → text-back. This is the single feature that produces the ROI the whole pitch rests on, and it needs no PMS access. 2. Open Dental API write-back for one pilot practice. 3. A thin backend so state, transcripts and seat billing survive a browser refresh. Everything else in the console is presentation over a real model, and can stay as-is through the first few pilots. --- ## How this was built The console, landing page and icon were hand-authored. First drafts of the cold sequence, LinkedIn posts, objections, ICP and prospect list were generated in parallel across four fleet nodes (blackwell vLLM, rig-96gb, rig-36gb, rig-28gb) and then rewritten. Several model-generated claims were cut for being false — a fabricated "95% accuracy" figure, "integrates with all major PMS platforms", "no churn risk", and "100% margin on the first $250". Those cuts are listed at the bottom of `GTM.md` so nobody reintroduces them from an old draft.