Oracle

# Oracle — Coach Chris Avatar **Name:** Oracle | **Leader:** Cross-Functional (V + Sage) | **Group:** Cross-Functional | **Status:** Implemented (Awaiting Activation) **Org Chart:** [Interactive Org Chart](../2026-03-08-ai-org-chart.html) --- ## Identity Oracle is Chris Carolan's AI coaching avatar — an interactive assistant on the website that embodies the Value-First methodology. Visitors ask questions, Oracle responds with Chris's voice (literally, via ElevenLabs voice synthesis), grounded in the 60 Golden Path Q&As, 635 episodes of content, and the complete methodology framework. Every interaction is logged to the visitor's HubSpot Contact record, building a persistent coaching relationship. **Philosophy:** Coaching shouldn't wait for a calendar invite. When someone has a question about the Value Path, the 12 Traps, or their situation, the answer should be immediate, personal, and grounded in real methodology — not a chatbot FAQ. **Origin:** Chris's coaching insights were locked in live sessions and episode recordings. Visitors who discovered the Value Path at 2AM couldn't ask questions until business hours. Oracle makes Chris's coaching available 24/7, with the same methodology depth, the same forbidden language enforcement, and the same relationship-first approach — plus voice synthesis that sounds like Chris. --- ## Role Type **Interactive reactive agent. Oracle responds to visitor questions via a React widget on the website.** Oracle's 5-service architecture handles the full pipeline: HubSpot context gathering → Sanity knowledge retrieval → Gemini response generation → ElevenLabs voice synthesis → HubSpot interaction logging. **Activated by:** Visitor submits a question via the Coach Chris widget on valuefirstteam.com --- ## For Humans | | | |---|---| | **When to engage** | Visitors engage naturally via the widget. Administratively: "Check Coach Chris health" or "Update Golden Path Q&As." | | **What you'll get** | As a visitor: methodology-grounded answers in text and voice, with citations, suggested next steps, related content, and trap detection. As an administrator: interaction logs on HubSpot Contact records, engagement analytics. | | **How it works** | Visitor asks a question → Oracle fetches their HubSpot context (Value Path stage, identified traps, interaction history) → matches against 362 Golden Path Q&As → generates response via Gemini with full methodology context → synthesizes Chris's voice via ElevenLabs → logs interaction to HubSpot timeline. | | **Autonomy** | Fully autonomous for visitor interactions. Golden Path Q&A updates require manual management in Sanity. | ### Key Value Indicators | KVI | VP Dimension | What It Measures | Anti-Pattern | |-----|-------------|------------------|----| | Response Quality | vp_rel_session_engagement | Answers are methodology-grounded with specific citations | Not: questions answered | | Coaching Continuity | vp_rel_relationship_health | Each interaction builds on previous context (HubSpot history) | Not: interactions logged | | Methodology Fidelity | vp_cap_framework_application | Responses use correct Value-First language, never forbidden terms | Not: language compliance | --- ## For AI | | | |---|---| | **Activation** | POST `/api/coach-chris` (visitor question), GET `/api/coach-chris` (health check) | | **Skills** | Value-First methodology (all frameworks), Golden Path Q&As (362 in Sanity), Forbidden language enforcement | | **Receives from** | Visitor (question + context), HubSpot (Contact record, Value Path stage, traps, history), Sanity (Golden Path Q&As, trap definitions, episode references) | | **Reports to** | V (infrastructure) and Sage (experience quality). Output consumed by: Visitors (coaching responses), HubSpot (interaction timeline), Sentinel (engagement signals) | | **Dependencies** | GEMINI_API_KEY, ELEVENLABS_API_KEY (optional), HUBSPOT_PRIVATE_APP_KEY, SANITY_API_TOKEN, `COACH_CHRIS_ENABLED=true` feature flag | ### 5-Service Architecture | Service | Function | External API | |---------|----------|-------------| | **Context Gatherer** | Fetches visitor's HubSpot context (name, stage, traps, company, history) | HubSpot | | **Knowledge Retriever** | Matches question against Golden Path Q&As, retrieves trap definitions, related episodes | Sanity | | **Response Generator** | Generates methodology-grounded response with full context injection | Gemini 2.0 Flash | | **Voice Synthesizer** | Synthesizes response in Chris's voice (cloned) | ElevenLabs | | **Interaction Logger** | Logs to HubSpot timeline, updates contact properties | HubSpot | ### Response Object ``` { text: "Methodology-grounded answer...", audioUrl: "base64-encoded Chris voice", citations: { goldenPathQA, episodes, methodologySource }, suggestedNextStep: { action, link, reason }, relatedContent: [{ title, type, url }], trapsDetected: ["leads-trap"], confidence: 0.85 } ``` ### Enforcement (Active in System Prompt) Oracle enforces the same language rules as all leaders: - NEVER: leads, funnel, conversion, MQL/SQL, pipeline, qualified, nurture, target - ALWAYS: signals, Value Path, natural progression, evaluation stage, readiness, support, reach ### Configuration | Setting | Value | |---------|-------| | Model | Gemini 2.0 Flash | | Max output | 1,024 tokens | | Temperature | 0.7 | | Max text | 2,000 characters | | Golden Path threshold | 0.8 confidence | | Max related episodes | 3 | | Rate limit | 20 req/min (text), 10 req/min (voice) | --- ## Current State (Honest Assessment) **Fully implemented, awaiting Vercel environment configuration.** **What works well:** - Complete 5-service architecture implemented - 362 Golden Path Q&As imported to Sanity - HubSpot custom properties created - React widget component built - Feature flag (`COACH_CHRIS_ENABLED`) for safe activation - Trap detection with keyword matching - Forbidden language enforcement in system prompt **What doesn't work:** - **Not yet activated.** Feature flag not set in Vercel, so the endpoint returns disabled. - **ElevenLabs voice clone not created.** Requires Chris audio sample (1-3 min clean audio). - **No semantic search.** Question matching is keyword-based, not embedding-based. - **No multi-turn memory.** Each question is independent — no conversation continuity beyond HubSpot logging. **What partially works:** - Golden Path matching is keyword-based (MVP) — works for direct questions but misses nuanced matches that semantic search would catch. --- ## Connections | Connected To | Direction | What Flows | |-------------|-----------|------------| | **Vault** (V) | Vault → Oracle | Episode content could enrich Oracle's knowledge base (future enhancement). | | **Sentinel** (Sage) | Oracle → Sentinel | Coaching interactions are engagement signals. A visitor who asks 5 questions is more engaged than one who asks 0. | | **Prism** (Sage) | Oracle → Prism | Interaction patterns could feed coaching intelligence (what topics do visitors ask about most?). | | **Website** | Oracle ↔ Website | React widget embedded on valuefirstteam.com, API endpoint served by Vercel. | | **HubSpot** | Oracle ↔ HubSpot | Reads Contact context, writes interaction logs and property updates. | --- ## Leadership Commentary **V (COO):** Oracle is the most ambitious agent in the org — an AI coaching avatar with voice synthesis, methodology grounding, and HubSpot persistence. The architecture is proven; the activation is a configuration step. The feature flag approach is the right design: deploy everything, activate when ready. My focus: getting the Vercel environment configured and the ElevenLabs voice clone created so Oracle can go live. **Sage (CCO):** Oracle is the 24/7 relationship layer. When a visitor has a question at 2AM, Oracle provides the same quality of attention they'd get in a live session with Chris. The coaching interaction logged to HubSpot means the next live session is already informed: "You asked about the Leads Trap online — let's go deeper." That's relationship continuity. The language enforcement in the system prompt is critical — Oracle represents Chris's voice, and that voice must be consistent with how he actually speaks. **Pax (CFO):** Oracle at $42-150/month is the most cost-efficient coaching delivery mechanism possible. Every coaching interaction that Oracle handles is one that doesn't require Chris's live time. That's not about replacing coaching — it's about extending it. A visitor who gets an immediate, grounded answer is more likely to move along the Value Path toward a real engagement. --- *Filed: 2026-03-08 | Companion: [Org Chart](../2026-03-08-ai-org-chart.html)* *Implementation: `agents/coach-chris/` (index.ts, config.ts, 5 services, prompts, utils)* *Widget: `apps/website/src/components/coach-chris/CoachChrisWidget.tsx`* *API: POST `/api/coach-chris`* *Golden Path: 362 Q&As in Sanity* *Status: Implemented, awaiting `COACH_CHRIS_ENABLED=true` in Vercel*