TruffleNation.ai — AI Pastry Chef Mentor App

01 Context & Vision

Product Description

TruffleNation.ai is an AI-powered pastry mentor available at trufflenation.ai. It delivers personalized pastry education through AI-generated courses, interactive quizzes, flash cards, audio lessons, and a comprehensive baker's toolkit — all grounded in a proprietary knowledge base built from real pastry expertise.

Proprietary Knowledge Base

29 TN class transcripts — Full day-by-day pastry school curriculum
~100 Scoolinary professional recipe books from 30+ international chefs — covering everything from viennoiserie to molecular pastry

Three Source Documents

Oboe Feature Docs
Pastry Chef Skill Explainer
TruffleNation.ai Product Outline

Ecosystem Integration

Student Portal at students.trufflenation.ai — shared authentication, same PostgreSQL database
82-tool Baker's Toolkit superapp — embedded directly into the AI Mentor platform

02 Tech Stack

Layer	Choice	Reasoning
Web	Next.js 14 (App Router)	Matches Student Portal stack; SSR for SEO; streaming API routes
Mobile	Expo (React Native)	Shares TypeScript/types; native feel; EAS for store submission
Backend	Next.js API routes on VPS (PM2 + Nginx)	Same infra as Student Portal; no timeout limits
Database	PostgreSQL 16 on VPS (existing)	Already running; add pgvector extension
ORM	Prisma	Already used in Student Portal
Vector DB	pgvector (PostgreSQL extension)	No external service; SQL-based filtering
Embeddings	OpenAI text-embedding-3-small	$0.02/M tokens; one-time ~$0.02
AI Generation	Claude Sonnet 4 (courses) + Haiku (quizzes, flash cards)	Sonnet for quality; Haiku 10-20x cheaper for lightweight
TTS	ElevenLabs	Two-voice conversational audio; $5/mo
Auth	JWT (jose + bcryptjs)	Same as Student Portal; shared DB
Payments	Razorpay (INR) + Stripe (USD)	Dual market; IP-based detection
File Storage	VPS filesystem → Cloudflare R2 (scale)	Start simple
Hosting	Hostinger VPS (existing)	$0 incremental; PM2 + Nginx
Email	Brevo	Existing TN integration
Monorepo	Turborepo	apps/web + apps/mobile + packages/shared

✓

VPS Consolidation Benefits
• $0 additional hosting cost
• Same PostgreSQL instance — shared users table
• No serverless timeout limits
• pgvector instead of Pinecone — one fewer external service
• Single deployment target: git pull → build → pm2 restart

Monorepo Structure

trufflenation-ai/
  apps/
    web/                 # Next.js 14 (trufflenation.ai)
    mobile/              # Expo (React Native, iOS + Android)
  packages/
    shared/              # TS types, API client, tier logic, baker's math, SR algorithm
    ui/                  # Design tokens (colors, spacing, fonts)

03 Database Schema

16 tables powering the entire AI Mentor platform. All tables live in the existing VPS PostgreSQL instance alongside the Student Portal.

id, email, password_hash, name, avatar_url, ai_tier (free/home_baker/chef), portal_student (boolean), credits_remaining, credits_reset_at, referral_code, referred_by, created_at, updated_at

id, user_id, title, topic, config_json (difficulty, style, focus), status (generating/complete/error), chapter_count, created_at, updated_at

id, course_id, order_index, title, content_md, rich_elements_json, sources_json, created_at

id, chapter_id, quiz_type (fill_blank/technique_match/fault_diagnosis), questions_json, created_at

id, quiz_id, user_id, answers_json, score, passed (boolean), attempted_at

id, user_id, course_id, cards_json, sm2_state_json (per-card interval, ease_factor, next_review), created_at, updated_at

id, course_id, user_id, content_json (6-section format), created_at

id, chapter_id, script_text, audio_url, duration_seconds, voice_config_json, status (pending/processing/ready/error), created_at

id, question_text, options_json, correct_answer, explanation, category, difficulty, publish_date, created_at

id, daily_bake_id, user_id, selected_answer, is_correct (boolean), responded_at

id, user_id, tier, payment_provider (razorpay/stripe/storekit), provider_subscription_id, status (active/cancelled/past_due), current_period_start, current_period_end, created_at

id, user_id, amount, action (course_generate/qa_query/quiz_generate/study_guide/flash_cards), reference_id, created_at

id, content, embedding (vector 1536), source_type, source_name, metadata (jsonb: day_number, instructor, topic_category, content_type, difficulty, section_title), created_at

id, user_id, bookmarkable_type (chapter/course/daily_bake), bookmarkable_id, created_at

id, user_id, course_id, last_chapter_index, progress_pct, last_accessed_at

id, user_id, code (unique), total_referrals, successful_referrals, current_tier, created_at

id, referrer_id, referee_id, referral_code, status (pending/verified/rewarded/rejected), verified_at, created_at

id, user_id, reward_type (extra_course/unlimited_chapters/flash_cards/study_tools/home_baker_month), activated_at, expires_at, created_at

04 RAG Pipeline

Chunking Strategy

Semantic boundaries — one technique, recipe, or concept per chunk
Target size: 300-500 tokens/chunk
Metadata: { day_number, instructor, topic_category, content_type, difficulty, source, section_title }
~2,000 chunks total

Retrieval Flow

Embed user query with text-embedding-3-small
Query pgvector with SQL WHERE for tier filtering
Re-rank by cosine threshold > 0.75
Assemble ~6,000 tokens of context
Pass to Claude with system prompt + retrieved context + user config

pgvector Setup

CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE knowledge_chunks (
  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  content TEXT NOT NULL,
  embedding VECTOR(1536),
  source_type VARCHAR(50) NOT NULL,
  source_name VARCHAR(200),
  metadata JSONB NOT NULL,
  created_at TIMESTAMP DEFAULT NOW()
);

CREATE INDEX ON knowledge_chunks
  USING ivfflat (embedding vector_cosine_ops)
  WITH (lists = 50);

System Prompt

You are the TruffleNation.ai Pastry Mentor. Draw EXCLUSIVELY from provided knowledge base.
Never fabricate techniques, temperatures, ratios, or chef quotes not in context.
Include: Chef's Corner (attributed quotes), technique tables, storage cards,
equipment callouts, baker's math, common mistakes, glossary terms [[in brackets]].

05 Phase 0 — Knowledge Base Preparation (Weeks 1-2)

Run all 29 TN class day transcripts through Whisper pipeline

Structure each transcript: techniques, recipes, theory, troubleshooting, ingredient science

Process Scoolinary content from SKILL.md (SC-1 to SC-29) into individual chunks

Install pgvector extension on VPS PostgreSQL

Generate embeddings (text-embedding-3-small) for all chunks

Insert into knowledge_chunks table with metadata

Validate: 50 test queries returning accurate, attributed results

06 Phase 1 — Core AI Platform (Weeks 3-6)

07 Phase 2 — Learning + Engagement + Referrals (Weeks 7-10)

Formula/ratio explainer with "Explain this" buttons + live calculators

In-lesson quizzes: fill-in-blank, technique match, fault diagnosis

Go Deeper: 5 AI follow-up questions + custom Q&A

Daily Bake: free daily trivia + email capture via Brevo

Social sharing: WhatsApp (primary) + Instagram + Facebook + Twitter

Course history sidebar + bookmarking

Sources panel: "View N Sources" with chef/book attribution

Referral system: unique codes, reward ladder (1→3→5→10), WhatsApp share, progress dashboard, anti-abuse

★

Demoable: Complete learning loop — generate, read, quiz, ask deeper, share, play Daily Bake.

08 Phase 3 — Study Tools + Audio + Payments (Weeks 11-15)

Flash cards: SM-2 spaced repetition, swipe UI

Study guide: 6-section format

Practical exam: 12 MCQ + gamification badge (80% pass)

Audio lessons: Claude two-voice script → ElevenLabs TTS → ffmpeg concat

3-tier monetization (Free / Home Baker ₹499 / Chef ₹1,499)

Razorpay (INR) + Stripe (USD) integration

Credit system with weekly resets

Portal student auto-upgrade to Chef tier

★

Demoable: Subscribe, listen to audio, study with flash cards, take exam, earn badge.

09 Phase 4 — Mobile Apps (Weeks 16-20)

Expo project setup with Expo Router

Shared package integration

Core screens: auth, course list, lesson viewer, flash cards, audio

Native features: expo-av, expo-notifications, expo-sqlite

Deep linking

Haptic feedback

EAS Build for TestFlight + Internal Testing

App Store + Play Store submission

10 Phase 5 — Polish + Launch (Weeks 21-24)

Portal auth linking

PDF export for Chef tier

Performance optimization

Error monitoring (Sentry)

Content pipeline automation

Admin dashboard

Launch email to 20K BakeBetterPro students

Cross-sell CTAs

11 Authentication

Web: JWT in httpOnly cookies
Mobile: JWT in expo-secure-store
Portal ↔ AI Mentor: Same user row, same JWT secret, ai_tier column
No cross-domain token dance needed — single domain, single database
OAuth (future): Google + Apple

ℹ

Because the AI Mentor and Student Portal share the same PostgreSQL instance and JWT secret, a user who logs in on one platform is effectively authenticated on both. The ai_tier column on the users table determines access level, and portal students are auto-upgraded to Chef tier.

12 Pricing & Monetization

Feature	Free	Home Baker (₹499/mo)	Chef (₹1,499/mo)
AI courses	1/mo, 2 chapters	10/week, unlimited	Unlimited
AI model	Sonnet	Sonnet	Sonnet
Glossary	Read-only	Full + interactive	Full + interactive
Go Deeper Q&A	3/month	50/day	Unlimited
Audio	No	Streaming	Streaming + download
Flash cards	No	Unlimited	Unlimited
Study guide	No	Yes	Yes
Practical exam	No	Yes	Yes + certificate
Baker's Toolkit	5 tools	All 82	All 82
Daily Bake	Yes	Yes + archive	Yes + archive
Scoolinary	No	No	Yes
PDF export	No	No	Yes
Public sharing	No	No	Yes

₹

Free tier cost: ~₹71/user/mo ($0.85). 1,000 free users = ~$850/mo.

13 Referral System — "Invite a Baker"

Rewards Ladder

Referrals	Unlock	What they get	Incremental cost
0	Base Free	1 course/mo, 2 chapters, 3 Q&A/mo	$0.85/mo
1	+1 course	2 courses/month	+$0.15/mo
3	+chapters +cards	Unlimited chapters + flash cards	+$0.05/mo
5	+study tools	Study guide + 10 Q&A/week	+$0.12/mo
10	Free Home Baker	Full tier for 1 month	~$2.40 one-time

✓

Economics: 10 referrals = $0.24 CAC per user vs $0.60-2.40 for Meta ads

Referral UX

Dashboard card with progress bar
WhatsApp share with pre-filled message
Push notifications on referral events
Anti-abuse: email verify + first course + 7-day active check

Viral Loop

Free user

→

Shares on WhatsApp

→

Friend installs

→

Signs up

→

Views course

→

Referrer gets reward

→

Shares more

↻

14 Infrastructure Costs

Monthly Infrastructure

Service	Monthly
Hostinger VPS	$0 incremental
PostgreSQL + pgvector	$0 incremental
EAS Free	$0
Claude API (cached)	~$150-250
ElevenLabs	$5
OpenAI Embeddings	~$1
Apple Developer	$8.25/mo
Total	~$165-265/mo

Claude API Breakdown

Task	Model	Per call	Monthly volume	Cost
Course generation	Sonnet	~$1.20	~100 (80% cache)	$120
Go Deeper Q&A	Sonnet	~$0.05	~2,000	$100
Quiz generation	Haiku	~$0.008	~500	$4
Study guide	Haiku	~$0.01	~300	$3
Flash cards	Haiku	~$0.005	~400	$2
Daily Bake	Haiku	~$0.003	30	$0.09

✓

Cost Reduction Strategies
1. Pre-generate top 50 courses (~$60 one-time)
2. Cache identical/similar topics
3. Haiku for lightweight tasks
4. Credit system limits free-tier generation

15 Break-Even Analysis

Scenario: 1,000 free users + 100 paying users

Monthly Costs

Item	Monthly
Claude API — paying users	~$229
Claude API — free users	~$850
ElevenLabs	$5
OpenAI	$1
Apple Dev	$8.25
Total	~$1,093/mo

Revenue Scenarios

Scenario	Revenue	Profit
100 Home Baker	~$600	-$493
50 HB + 50 Chef	~$1,200	+$107
100 HB + 20 Chef	~$960	-$133
200 Home Baker	~$1,200	+$107

⚠

Key insight: Need ~170 Home Baker users OR healthy HB+Chef mix to cover 1,000 free users.

ℹ

At 500 free users: Total cost ~$668/mo. Break-even at 112 Home Baker or 56 mixed.

16 Risks & Mitigations

Risk	Severity	Mitigation
RAG wrong pastry info	HIGH	Prototype in Phase 0; 50 test queries; low temp (0.3); source attribution
Claude API cost spike	MEDIUM	Credit system; rate limiting; email-required registration
Apple rejects Razorpay	MEDIUM	StoreKit for iOS; Razorpay Android+web only
KB incomplete	MEDIUM	Phase 0 dedicated; batch Whisper; Scoolinary already comprehensive
Timeline too aggressive	MEDIUM	MVP ships in 6 weeks; validate before committing remaining 18

17 Launch Strategy

Week	Milestone
6	Closed Alpha — 50 hand-picked students
10	Open Beta — Email 20K students; target 2K signups
15	Soft Launch — Payments enabled; beta users get 1 month free
20	Mobile Launch — App Store + Play Store
24	Full Launch — PR, referral program, Meta ads

18 Verification Plan

Run 50 diverse test queries across all 29 class days and Scoolinary sources
Verify chef attribution is correct (no misattributed quotes)
Check cosine similarity scores — ensure top results exceed 0.75 threshold
Test tier-based filtering — free users should not retrieve Scoolinary chunks
Validate chunk boundaries — no mid-sentence splits or orphaned technique steps
Confirm embedding count matches expected ~2,000 chunks

End-to-end course generation: topic input → streaming output → rendered chapters
Auth flow: register → login → JWT validation → protected routes
Rich renderer: all 6 component types display correctly (ChefTip, TechniqueTable, StorageCard, EquipmentCallout, BakersMathBlock, ChefsCornerSidebar)
Glossary: click term → side panel opens with correct definition
Baker's Toolkit: all 82 tools load in iframe/embed correctly
Mobile responsive: test on iPhone SE, iPhone 14 Pro, Pixel 7
Streaming performance: first token under 1s, full course under 30s

Quiz generation: all 3 types produce valid, answerable questions
Go Deeper: AI follow-ups are contextually relevant to current chapter
Daily Bake: new question appears each day; email capture triggers Brevo flow
Social sharing: WhatsApp deep link opens correctly with pre-filled message
Bookmarking: save, retrieve, and remove bookmarks across sessions
Sources panel: displays correct source count and attribution
Referral system: code generation, tracking, reward unlocks at correct thresholds
Anti-abuse: verify email requirement, first course completion, 7-day active check

Flash cards: SM-2 algorithm produces correct review intervals
Study guide: all 6 sections generate with relevant content
Practical exam: 12 questions, scoring, 80% pass threshold, badge award
Audio: script generation → TTS → concat produces playable audio
Razorpay: test payment in INR → subscription activates → tier upgrades
Stripe: test payment in USD → subscription activates → tier upgrades
Credit system: weekly reset fires correctly; limits enforced
Portal student: auto-detects portal status → upgrades to Chef tier

iOS: all core screens render correctly on iPhone SE through iPhone 15 Pro Max
Android: test on Pixel 7, Samsung Galaxy S23, budget device (e.g., Redmi)
Audio playback: background audio, lock screen controls, notification
Push notifications: daily bake reminder, referral event, new course ready
Deep linking: web URL opens correct screen in app
Offline: SQLite caches last-viewed course for offline reading
TestFlight: internal testers can install and use full flow
Play Store internal testing: APK installs and runs without crashes

Portal auth: student portal user can log into AI Mentor without re-registering
PDF export: Chef tier user exports course → formatted PDF downloads
Performance: Lighthouse score > 90 for performance, accessibility, SEO
Sentry: errors are captured and reported with correct source maps
Content pipeline: new transcript → auto-chunk → embed → available in RAG
Admin dashboard: user count, revenue, popular topics, error rate visible
Launch email: Brevo campaign sends to 20K list with correct links and tracking
Cross-sell: AI Mentor → Portal CTAs render; Portal → AI Mentor CTAs render
Load test: simulate 100 concurrent users generating courses without errors