AI & ML Overview
Document Purpose: This document defines the Pragmatic AI strategy for Cazo. We reject "AI-washing" in favor of a cost-effective, tiered approach: Rules First, Small Models Second, and LLMs Last. MVP focus is on Slot Optimization and AI Nudges.
Executive Summary
Cazo AI transforms "Empty Chairs" into "Revenue". Our MVP AI strategy is laser-focused on maximizing seat occupancy through intelligent nudges and slot management. Futuristic features like Computer Vision are explicitly deferred to Phase 2+.
Stakeholder Directive (Dec 2024): "AI Nudges to improve seat occupancy is the most critical bit... Slot management and AI nudges for improving sales should be the part of first build."
1. MVP AI Focus: "Empty Chair Rescue"
The core AI value proposition for the first build.
1.1 Slot Management (Deterministic + Light AI)
| ID | Capability | Implementation | Value |
|---|---|---|---|
[SLOT-001] |
Gap Detection | n8n Rule: "Find slots > 30 mins empty." | Identify revenue leakage. |
[SLOT-002] |
Smart Rebalancing | Suggest moving booking to fill gap. | Improve utilization %. |
[SLOT-003] |
Processing Time Awareness | "Color + 45m processing" auto-reserves. | Prevent accidental overbooking. |
1.2 AI Nudges (The Revenue Engine)
| ID | Capability | Implementation | Value |
|---|---|---|---|
[NUDGE-001] |
Empty Chair Alert | "3 empty slots tomorrow PM. Target lapsed clients?" | Fill gaps proactively. |
[NUDGE-002] |
Re-booking Prompt | "Jane (Color) visited 6 weeks ago. Send nudge?" | Increase repeat visits. |
[NUDGE-003] |
Upsell Suggestion | "Client booked Cut. Suggest Repair Treatment?" | Increase ticket value. |
2. The Pragmatic Intelligence Hierarchy
We process all requests through a strict funnel to minimize cost and hallucination risk.
| Tier | Technology | Use Case | Status |
|---|---|---|---|
| L1: Deterministic | n8n / SQL | Slot gap detection, Booking confirmation. | MVP |
| L2: Specialized SLM | Mistral 7B | Intent extraction from WhatsApp. | MVP |
| L3: Generalist LLM | GPT-4o (API) | Complex empathy, Angry customer handling. | MVP (Fallback) |
3. Deferred to Phase 2+ (Futuristic)
These capabilities are valuable but not in the first build.
| ID | Capability | Reason for Deferral |
|---|---|---|
[VIS-001] |
Face Shape Analysis | Requires camera integration, model training, high R&D. |
[VIS-002] |
Hair Condition Grading | Needs high-res image pipeline, edge deployment. |
[VIS-003] |
Inventory Scanning | Barcode scanning is sufficient for MVP. |
4. Marketing Intelligence (Phase 2)
Advanced AI capabilities for the Marketing Studio, to be built after MVP stabilization.
4.1 Campaign Analytics & Summarization
| ID | Capability | Description |
|---|---|---|
[MKTG-001] |
Campaign Summarization | "Summarize last month's campaign performance in 3 bullets." |
[MKTG-003] |
A/B Test Analysis | "Which message variant performed better and why?" |
4.2 AI-Powered Segmentation
| ID | Capability | Description |
|---|---|---|
[SEG-001] |
RFM Clustering | Recency / Frequency / Monetary scoring to auto-classify clients (VIP, At-Risk, Lost). |
[SEG-002] |
Lookalike Audiences | "Find 50 clients most similar to my top 10 spenders." |
[SEG-003] |
Churn Prediction | ML model to score each client's likelihood to lapse within 30 days. |
[SEG-004] |
Upsell Propensity | "Which clients are most likely to accept a Treatment add-on?" |
4.2 Self-Orchestrating Agent (Future Vision)
When marketing requires autonomous, multi-step reasoning, we will consider an agent framework (LangGraph / CrewAI).
- Trigger: "Run marketing on autopilot."
- Agent Loop: Analyze Data -> Select Audience -> Draft Message -> Schedule Send -> Measure -> Iterate.
- Pre-requisite: Trust built through successful Phase 1 nudges with human approval.
5. Inference Infrastructure
4.1 Self-Hosted Core ("The Private Cloud")
For high-volume, privacy-sensitive workflows.
- Tech Stack: Ollama on AWS EC2
g5.xlarge. - Models:
Mistral 7Bfor intent parsing.
4.2 Public API ("The Genius Cloud")
For complex reasoning tasks only.
- Tech Stack: OpenAI API.
- Models:
GPT-4o. - Usage: <5% of total request volume.