BWH Business
Operating System

Implementation Plan

Prepared for Adrian Waldron & Reggie Black
Date March 10, 2026
Version 1.0
Update · April 28, 2026
Phase 2 findings from 12 months of Zendesk data are now available — five analysis passes, recommended A/B test with real baselines.
Read the Findings →
01

Executive Summary

Better Way Health's data lives across 18+ platforms. No single person can answer cross-functional questions without pulling reports from multiple tools. ScrappyLabs will build a modular, AI-powered Business Operating System that unifies everything into one intelligent layer.

4
Core Modules
18+
Data Sources
12
Week MVP
6
AI Agents

The core thesis: One unified data layer. Multiple specialized AI agents drawing from it. A Progressive Web App as the interface. Built on Supabase + FastAPI + React PWA with a model-agnostic AI layer (Claude Opus, GPT-5.2, Haiku — selectable per use case).

12-week MVP timeline with milestone-based delivery.

02

What We're Building

Four modules, each drawing from the same unified data layer. Modules 1–3 ship in the 12-week MVP. Module 4 follows as a separate engagement.

MVP — Module 1
≡
RAG Chat Agent
Ask anything about the business
  • 6 specialized sub-agents: Research, Sales, CS, Marketing, Operations, Webinar
  • Orchestrator classifies intent and routes to the right agent(s)
  • Every factual claim cited with expandable source modals
  • Multi-agent queries combine insights across domains
  • Read-only — retrieves and synthesizes, never takes actions
MVP — Module 2
◇
Webinar & KOL Management
Plan, track, optimize
  • Calendar view with drag-and-drop scheduling
  • 130-task template auto-generated per webinar
  • KOL CRM: profiles, performance history, card grid
  • Revenue attribution: discount code → Shopify → webinar
  • Comparison views: KOL vs KOL, webinar vs webinar, trends
MVP — Module 3
✎
Copywriting Engine
35+ assets per webinar, AI-generated
  • Emails, SMS, landing pages, sales pages, scripts — all from context
  • Pulls from webinar record + KOL profile + research + product data
  • Prompt management studio with template variables & version history
  • Single-layer approval workflow (CMO reviews)
  • Per-asset model selection (Opus for creative, Haiku for fast)
Post-MVP — Module 4
☎
Voice Sales Agent
B2B outbound to 8,000 HCP leads
  • AI outbound calls — 20 calls/day target
  • Dynamic scripting, objection handling, CRM integration
  • TCPA compliance review required before build
  • ScrappyLabs has production voice agent infrastructure already in use
  • Separate SOW after MVP is stable
03

Architecture

Clean separation of concerns. The frontend talks to one API gateway. The gateway routes to agents or modules. Everything reads from and writes to Supabase.

PWA Frontend
React · Installable · Mobile-optimized · Real-time via Supabase
↓↑
FastAPI Gateway
Auth middleware · Rate limiting · Request routing · Logging
↓↑
Agent Orchestrator
Intent classification · Permission check · Multi-agent dispatch
Research
Sales
CS
Marketing
Operations
Webinar
↓↑
Data Connectors
Unified pattern · Incremental sync · Webhook + batch
Zendesk Shopify Recharge Fathom Klaviyo Zoom PubMed Everflow Notion Slack Drive
↓↑
Supabase
PostgreSQL · pgvector · Auth (Google OAuth) · RLS · Realtime
04

Data Connectors

Every data source follows the same four-method pattern: Configure, Validate, Schema, Sync. Build the pipeline once, add connectors by filling in the template.

1. Configure
What credentials we need
2. Validate
Test the API connection
3. Schema
Available tables & resources
4. Sync
Incremental + full refresh
Source Priority Sync Mode Auth Volume
Zendesk (tickets) HIGH Near-RT webhook API token 600+/day
Fathom HIGH Near-RT webhook API key 150 calls/week
PubMed HIGH On-demand Free 1,000+ docs
Shopify HIGH Daily batch Access token 5K orders/mo
Recharge HIGH Daily batch API token 9K subs
Zoom MEDIUM Post-event Server OAuth 4/month
Klaviyo MEDIUM Daily batch API key 100K list
Everflow LOW Daily batch API key 300 affiliates
Amazon SP-API LOW (defer) Daily batch Complex 11 ASINs
First connector takes 2–3 days to build. Each subsequent connector: 1–2 days (shared pipeline).
05

Timeline

Twelve weeks from start to soft launch. Three phases, each with a milestone payment. Weekly demos every Friday.

Phase 1: Foundation
Weeks 1–4 · April 2026
Week 1 Supabase project + schema, Google OAuth, GCP Cloud Run, PWA shell
Week 2 Data pipeline architecture, Zendesk + Fathom + PubMed connectors
Week 3 RAG pipeline: chunking strategies, embedding, pgvector search
Week 4 Chat MVP: orchestrator, Research + CS + Webinar agents, citations UI
Phase 2: Core Modules
Weeks 5–8 · May 2026
Week 5 Webinar calendar + task management + Zoom API integration
Week 6 KOL CRM: profiles, performance tracking, card grid with sorting
Week 7 Analytics dashboard + Shopify discount code → webinar attribution
Week 8 Remaining connectors: Shopify, Recharge, Klaviyo, Everflow
Phase 3: Copywriting & Polish
Weeks 9–12 · June 2026
Week 9 Copywriting engine + context pipeline (webinar + KOL + research)
Week 10 Prompt management studio + approval workflow
Week 11 Admin panel: user management, permissions grid, remaining agents
Week 12 QA, performance tuning, mobile optimization, team training & launch
Phase 4: Voice Agent
Q3 2026 · Separate SOW
Post-MVP B2B voice sales agent, compliance review, call analytics, CRM integration
06

What ScrappyLabs Brings

We're not starting from zero. Every major component BWH needs maps to production infrastructure we already operate.

BWH Needs
We Already Have
RAG system for 1,000+ research docs
Production RAG pipeline with vector search, running today
Multi-agent chat orchestration
Agent orchestrator patterns, 6+ AI agents in production
Voice sales agent (Phase 4)
Production voice agent with telephony + TTS, already deployed
Real-time data synchronization
Production WebSocket and real-time data relay systems, battle-tested
Supabase + vector search
Multiple Supabase projects with RLS, pgvector, auth
AI model flexibility
6 AI coding CLIs, local GPU fleet, all major cloud APIs
Apprenticeship / mentoring model
Built an entire fleet teaching lab — mentoring is the culture
07

Development Approach

Claude Code First
Primary development tool, as specified in the PRD. AI-accelerated development from Day 1.
Multi-Agent Dev
6 AI coding CLIs working in parallel on independent modules. Faster delivery, same quality.
Collaborative Build
Architecture walkthroughs, code reviews, and knowledge transfer throughout the project.
Weekly Demos
Every Friday: show what's working, get feedback, adjust priorities for the next week.
GitHub Repo
scrappylabsai/bwh-bos (private). Adrian gets access Day 1.
Communication
Slack channel or your preferred tool. Async updates + weekly syncs. No surprises.
08

What We Need from BWH

To start building on Day 1, we'll need the following. Sandbox or test credentials are preferred — we can work with production access if that's easier.