Stewart AI - System Architecture

1

Business Entities

🏋

Ark Gym

Dubai. Launch August/September 2026. Member management via Mindbody. SOPs for operations, pricing, membership rules, staff policies. Knowledge isolated to gym-authorized staff.

🎥

LiveWire Productions

Digital content and production. Customers: gym equipment manufacturers and gyms globally. AI lip sync, set building, logo integration, video production. Six years of content, three years as LiveWire.

🏢

Stewart Miller Company

Umbrella entity. Cross-company policies, financial data, strategic plans, C-suite intelligence. Only accessible to senior management and above.

2

Knowledge Pipeline - How Information Enters the System

OneDrive

SOPs Created

→

Google Drive

Synced Copy

→

EA Validates

Stewart Approves

→

Stewart OS
Processed & Indexed

→

Team Access

Role-Filtered

Needs Work

📄

SOP Ingestion

New SOPs, company policies (visa, leave, pricing, membership rules, new staff). EA uploads daily after Stewart validates. BuddyPro uses Google Drive for ingestion: EA uploads files to a shared Drive folder, then runs /update in Telegram. No API for programmatic ingestion. Custom build can auto-detect new files and process them without manual steps.

Custom Build

📊

Mindbody BI Data

Member counts, retention rates, class attendance, revenue metrics. Requires a data integration to pull live data into the AI. BuddyPro's API is chat-only (no external data connectors). A custom data pipeline is needed to bridge Mindbody into the knowledge layer.

BuddyPro OK

📚

Static Knowledge Upload

PDFs, documents, training materials, recorded content, YouTube/Vimeo URLs. BuddyPro handles this well via Google Drive folder + /update command. Supports PDF, docx, txt, audio, video, images, Google Docs, and web URLs.

3

Access Control - Who Sees What

Custom Build

BuddyPro Cannot Do This

BuddyPro has a 3-tier role system (Owner, Team, End User) but this controls admin commands, not knowledge access. All users see the same knowledge base. There is no per-user knowledge filtering. To simulate tiered access, you would need multiple separate BuddyPro instances ($197/month each, plus per-message AI costs) with no unified management view. Even then, conditional visibility within an entity (management sees more than team) is not possible.

Person	Role	Ark Gym	LiveWire	Stewart Miller Co.	Financial Data	All SOPs
Stewart CEO / Owner	C-Suite	✓	✓	✓	✓	✓
Thushani Executive Assistant	Senior Mgmt	✓	✓	✓	✓	✓
Kim Project Manager	Senior Mgmt	✓	✓	✓	Ltd	✓
Dragon Gym Operations	Management	✓	Pub	—	—	—
Chris LiveWire Production	Management	Pub	✓	—	—	—
Gym Staff Team Members	Team	Ltd	Pub	—	—	—
Public / Clients Future Phase	Public	Pub	Pub	—	—	—

✓ Full Access Ltd Limited / Role-Specific Pub Public Info Only — No Access

4

Interface Layer - How People Access Stewart AI

BuddyPro OK

✈

Telegram Bot

Current interface. Team messages Stewart AI via Telegram. BuddyPro supports this today. Works for basic Q&A but lacks role-based filtering per user.

Custom Build

🌐

Web Dashboard

Admin view for Stewart and senior management. User management, knowledge base health, usage analytics, access control configuration. BuddyPro's analytics are accessed via Telegram /stats command (not a web dashboard). No visual admin interface for managing users or access levels.

Partially Possible

📱

Custom App (Matthew)

Stewart's son Matthew is building a custom app. BuddyPro does have an OpenAI-compatible chat API (api.buddypro.ai). Matthew's app could call it for chat. However: the API is stateful (all calls save to memory), has no multi-user support (no per-user identity), and the owner can read all test profile conversations. A custom API layer gives full control over user identity and privacy.

BuddyPro Possible

🧠

Public AI Coach

Future: paid access to Stewart's brain for clients and the public. This IS what BuddyPro was designed for. Could remain on BuddyPro for this specific use case while internal system is custom.

5

BuddyPro Capability Assessment

Requirement vs. BuddyPro Capability

Static Knowledge Upload

Upload PDFs, videos, documents to create an AI knowledge base. BuddyPro handles this well.

Telegram Chat Interface

Team can message the AI via Telegram. Currently working.

Public Coaching Product (Future)

Clients pay to access Stewart's brain. This is BuddyPro's core use case. Good fit for Phase 3.

Daily SOP Updates

EA uploads files to a shared Google Drive folder, then runs /update in Telegram. No API for programmatic ingestion, no file-watch capability, no automation. Works at small scale, becomes a daily manual task as document volume grows.

Role-Based Access Control

BuddyPro shows all knowledge to all users. No way to restrict Dragon to Ark only or Chris to LiveWire only. Would require separate instances ($197/mo each) with no unified view.

Multi-Entity Knowledge Separation

One knowledge base per BuddyPro instance. Cannot show management-level data to managers while showing public-level data to team within the same system.

Mindbody / External Data Integration

BuddyPro's API is chat-only. No data connector endpoints exist. Knowledge is added exclusively via Google Drive + /update command. Cannot pull live data from Mindbody or any external system.

Stewart Miller Branding

BuddyPro offers white-label at the Telegram bot level: you set your own bot name, avatar, and description. End users don't see "BuddyPro" in conversation. However, the delivery channel is still Telegram (not a Stewart-branded app). A planned native mobile app (targeted Q1 2026) has not yet shipped.

Custom App API

BuddyPro has an OpenAI-compatible chat API that Matthew's app could call. However, it is stateful (all calls save to memory), has no per-user identity field, no multi-user support, and the instance owner can read all test profile data. Works for basic chat, insufficient for a multi-user app with privacy isolation.

Coach Cloning / AI Avatars

BuddyPro does offer voice cloning (clone your voice for AI audio responses) and text-based AI experts. However, Stewart's vision of filming coaches once and creating video avatars in an app requires video generation (HeyGen-style), not text/audio chat. Different product category for the full vision.

Data Ownership and Portability

Source content stays in your Google Drive (you own those files). However, BuddyPro's processed knowledge base, user memory profiles, conversation histories, and subscription data live on their servers with no documented export. If BuddyPro shuts down, you keep raw files but lose all AI processing, memory, and user data.

6

Implementation Phases

Phase 1 - The Foundation

Stewart OS: Internal Operations

Custom AI brain with knowledge ingestion pipeline
Role-based access control (all team tiers)
Multi-entity knowledge separation (Ark, LiveWire, SMC)
Telegram bot with per-user filtering
SOP upload and processing workflow
Team training (waves: EA + PM, then managers, then staff)

Months 1-3 (90-day window, before gym launch)

Phase 2 - Expansion

Integrations and Intelligence

Mindbody API integration (live gym BI data)
Web dashboard for admin and analytics
API layer for Matthew's app
Chris / LiveWire AI production workflow
Advanced SOP automation (auto-detect new files)

Months 4-6

Phase 3 - Monetization

Public AI Coach and Beyond

Public-facing "Stewart's Brain" coaching product
Coach cloning infrastructure (AI avatars + voice)
Integration with Matthew's app marketplace
Client payment and subscription management
Scaling to Stewart's network of 1000+ coaches

Months 6-12