DocsBERTHA Development Summary

BERTHA Development Summary

BERTHA Development Summary

Overview

BERTHA is an AI collection intelligence agent trained by Amanda Schmitt. This document summarizes the development of BERTHA's framework, Claude SDK integration, and trainer interview system.

Architecture Components

1. Claude SDK Integration

Location: /src/lib/agents/bertha/claude-sdk.ts
  • • Bridges Claude AI capabilities with Eden infrastructure
  • • Key methods:
    • - analyzeOpportunity(): Market analysis for assets - generateStrategy(): Portfolio optimization - processTrainerInterview(): Training data processing

    2. Trainer Interview System

    Location: /src/app/sites/bertha/interview/page.tsx
  • • Streamlined from 30+ generic questions to ~20 psychology-focused questions
  • • 6 sections capturing collection decision-making:
    • 1. First Instincts (3 questions) 2. Emotion vs Logic (3 questions) 3. Risk & Timing (4 questions) 4. Social Dynamics (3 questions) 5. Vision & Value (4 questions) 6. Beyond Safety (3 questions)
  • • Includes trainer identification (name/email)
  • • No popup alerts - elegant inline feedback

    • 3. Amanda Bootstrap Knowledge

    Location: /src/lib/agents/bertha/amanda-bootstrap.ts
  • • Pre-loaded expertise from Amanda Schmitt
  • • Includes:
    • - Collecting principles and taste profile - Market intelligence and pricing strategies - Platform preferences and timing patterns - Red flags and evaluation framework

    4. Training Data Storage

    Location: /src/lib/agents/bertha/training-storage.ts
  • • JSON file-based storage (serverless-compatible)
  • • Non-blocking operations with fallbacks
  • • CSV export functionality
  • • Email notification system (placeholder)

    • 5. API Endpoint

    Location: /src/app/api/agents/bertha/training/route.ts
  • • POST: Process training submissions
  • • GET: Retrieve training status
  • • Enhanced error handling and logging
  • • Serverless-optimized with graceful fallbacks

    • 6. Admin Dashboard

    Location: /src/app/admin/bertha-training/page.tsx
  • • View all training submissions
  • • Display trainer info and responses
  • • Export capabilities
  • • Real-time status monitoring

    • Key Design Decisions

    1. Psychology-First Approach

    Instead of generic preference checkboxes, the interview captures the psychological framework behind collection decisions. This provides richer training data for BERTHA to understand Amanda's intuitive decision-making process.

    2. Serverless Compatibility

    All file operations are non-blocking with fallbacks, ensuring the system works in Vercel's serverless environment even if file storage fails.

    3. No Popups Policy

    Following user preference, all user feedback is inline with elegant state management rather than jarring alert() popups.

    4. Branding Clarity

  • • BERTHA: The AI agent (what users interact with)
  • • Amanda Schmitt: The human trainer (who teaches BERTHA)
  • • Academy page shows BERTHA as the agent with Amanda as trainer

    • Data Flow

  • Trainer fills interview → Name/email + 6 sections of questions
  • Form submission → Client-side validation → API endpoint
  • API processing → Claude SDK analysis → Storage attempts
  • Storage layers → Primary JSON file → Fallback logging → Optional Google Sheets
  • Admin monitoring → Dashboard at /admin/bertha-training

    • Outstanding Items

    To Fix

  • • URL structure: /sites/amanda should become /sites/bertha (noted for future)

    • Completed

  • • ✅ Claude SDK integration
  • • ✅ Comprehensive interview form
  • • ✅ Amanda bootstrap knowledge
  • • ✅ Training data persistence
  • • ✅ Admin dashboard
  • • ✅ Trainer identification fields
  • • ✅ Removed works section from Amanda's site
  • • ✅ Fixed BERTHA branding in Academy

    • Testing the System

    Interview Form

    Visit: /sites/bertha/interview
  • • Fill in trainer name and email
  • • Answer questions across 6 sections
  • • Submit to process training data

    • Admin Dashboard

    Visit: /admin/bertha-training
  • • View all submissions
  • • Check trainer details
  • • Review responses
  • • Export data

    • API Status

    GET: /api/agents/bertha/training
  • • Returns current training status
  • • Shows capabilities and readiness

    • Technical Stack

  • • Frontend: Next.js, React, TypeScript
  • • Styling: Tailwind CSS
  • • API: Next.js API Routes
  • • Storage: JSON files (development), Database-ready
  • • AI Integration: Claude SDK
  • • Deployment: Vercel

    • Security Considerations

  • • Input validation on all form fields
  • • Sanitized data before storage
  • • API token authentication (prepared)
  • • No sensitive data in client-side code
  • • Trainer emails stored but not exposed

    • Next Steps for Production

  • Replace file storage with database (Supabase/PostgreSQL)
  • Implement real email notifications
  • Add authentication to admin dashboard
  • Set up Google Sheets integration
  • Complete URL restructuring (amanda → bertha)
  • Add rate limiting to API endpoints
  • Implement training data versioning