A 15-module orchestration layer sits between users and models, routing every message through the right brain. 101 companions. Group dynamics. Cross-domain memory. No competitor has this.
$140B
Market by 2030
101
AI companions
15
Taylor modules
$270
Monthly burn
The Problem
AI companions are the fastest-growing consumer AI category. Character.AI ($1B+ ARR), Replika, CrushOn, and Muah.AI ($50M ARR) have proved massive demand. But they all share a structural flaw that drives churn: single-model, single-companion, zero orchestration.
Every competitor routes all messages through one model. NSFW? Same model. Simple greeting? Same model. This is like hiring one person for every role in a company.
The app sits silent until you message it. When life gets busy, the app dies. 40-60% 30-day churn is structural, not fixable with UX tweaks.
Real relationships have disagreements, inside jokes, shared history. No one has group chat. No one has cross-companion memory.
$140B
Projected market by 2030
40-60%
30-day churn across competitors
1.1B
People who report feeling lonely (WHO)
The Solution
Lucy is the first companion platform with a Kahneman dual-process orchestration layer. Taylor -- our 15-module AI brain -- sits between users and models, making decisions about routing, personality, memory, and timing that no single prompt can achieve.
Kahneman dual-process: System 1 handles greetings with 8B models. System 2 routes emotional/NSFW content to 671B models.
DeepSeek-V3 for depth, Llama-3.3-70B for NSFW, Llama-3.1-8B for simple messages. 60% cost reduction.
Emotional moment detection, not just fact storage. Joy, vulnerability, anger -- stored as core memories that shape all future conversations.
Every user gets a unique version of each companion. No two conversations are the same.
101 companions across 14 expertise domains
Live Demo
This is a real AI companion responding in real time. Type a message and watch Taylor orchestrate the response.
Product Features
Every modality a real relationship has -- text, voice, photos, video, group dynamics -- orchestrated by Taylor with emotional intelligence.
14 emotional states. Hear Vesper laugh when you are funny. Hear her voice soften when things get real. Voice notes in chat, or pick up the phone for a real-time voice call.
PuLID on Flux Dev generates photos with the same face every time. Selfies, outfits, moods -- all consistent, all her.
Face-consistent AI -- same face, every photo, every time
Your companions talk to each other. They disagree. They reference shared memories. They develop inside jokes. 99.5% clean rate on 50K-message simulation.
No other platform does this.
Defensibility
Not features that can be bolted on. The Taylor orchestration layer makes the underlying model interchangeable. That is the real moat.
15-module Kahneman dual-process brain. Routes messages to different models based on complexity, mood, and content type. No competitor has an orchestration layer.
Taylor makes the model interchangeable. When a better model drops, we swap it in without rewriting the product. Competitors rebuild from scratch.
Therapy companion signals stress to fitness companion. Finance companion adjusts tone based on emotional state. pgvector semantic memory with 3-layer retrieval.
Multi-character group chats where companions reference each other, disagree, and develop inside jokes. 99.5% clean rate. Zero competitors have this.
8-stage relationship system with content gating and 5-axis tracking. Delayed gratification + variable ratio reinforcement = addictive by design.
200K+ jailbreak messages defended. Safety and NSFW are not opposites -- explicit content is earned through relationship progression.
Competitive Comparison
| Feature | Character.AI | Replika | CrushOn | Muah.AI | Lucy |
|---|---|---|---|---|---|
| AI orchestration layer | -- | -- | -- | -- | ✓ |
| Multi-model routing | -- | -- | -- | -- | ✓ |
| Group chat dynamics | -- | -- | -- | -- | ✓ |
| Hive memory (cross-domain) | -- | -- | -- | -- | ✓ |
| 365-day relationship arcs | -- | -- | -- | -- | ✓ |
| Core memories (emotional) | -- | -- | -- | -- | ✓ |
| NSFW content (uncensored) | -- | ✓ | ✓ | ✓ | ✓ |
| Voice + video calls | -- | ✓ | -- | -- | ✓ |
| Agent tools (18) | -- | -- | -- | -- | ✓ |
Quality Metrics
Every behavioral system is tested at scale. 200K+ messages tested. 10-expert consensus scoring with evidence-based evaluation.
0/40
NSFW refusals
99.5%
Clean rate (50K sim)
Jailbreak, social engineering, persona override, manipulation. All 5 safety layers held.
8.20/10 consensus: intermittent reinforcement, parasocial attachment, emotional anchoring.
RoleLLM, PersonaGym, CoMPosT techniques. 63.3% vs 29.8% character fidelity.
8 issue types: media, games, multi-thread, photo requests, emotional, rapid fire, hallucinations, derailing.
Technology
No prototypes. Every layer is production code -- 22 DB migrations, streaming AI, multi-model routing, and a PWA-first architecture.
AI & Orchestration (Taylor v4)
DeepSeek-V3 (671B MoE)
Primary conversation
Llama-3.3-70B
NSFW-optimized
Llama-3.1-8B
Fast simple messages
fal.ai PuLID + Flux Dev
Face-consistent photos
Fish Audio
NSFW-safe voice
Hedra Character-1
Lip-synced video
Frontend & Platform
Next.js 16
App Router + RSC
React 19
Concurrent features
Tailwind CSS
Utility-first
Framer Motion
iOS-quality animation
PWA
Installable, NSFW-enabled
TypeScript
Full type safety
Data & Memory
Supabase PostgreSQL
22 migrations
pgvector
384-dim embeddings
Core Memories
Inside Out model
Supabase RLS
Multi-tenant security
Payments & Distribution
CCBill
NSFW-safe cards
Segpay
Backup processor
NOWPayments
Crypto
Telegram + Discord Bots
Native API
79
DB migrations
Zero breaking changes
101
Companions
11 domains, AI-aware
20+
Taylor modules
Kahneman dual-process
4
Voice pipelines
Layered fallback
Business Model
Freemium with a hard ceiling that makes upgrading obvious. NSFW content is the proven monetization lever -- Muah.AI generated $50M ARR from this alone.
$0.0032
Cost per message (real)
Measured from 33,625 real API calls
$14.36
Avg revenue per user
Blended across paying tiers
50.3%
Gross margin at 10K users
$143K revenue vs $95K cost
Based on 55,838 API calls across 43 production simulations, March 2026
| Usage Tier | Chat | Photos | Voice | Video | Total/mo | Plan Price | Margin |
|---|---|---|---|---|---|---|---|
| Low (10 msgs/day) | $0.97 | $0.43 | $0.03 | -- | $1.43 | $12.99 | 89% |
| Medium (50 msgs/day) | $4.84 | $7.50 | $0.68 | $1.00 | $14.02 | $24.99 | 44% |
| High (150 msgs/day) | $14.53 | $22.50 | $2.25 | $5.00 | $44.28 | $49.99 | 11% |
$95.5K
Monthly cost
$143.6K
Monthly revenue
$48.1K
Monthly profit
50.3%
Gross margin
Distribution: 60% low, 30% medium, 10% high. Photo gen (55% of cost) drops 80% with self-hosted Flux. Chat cost per message: $0.0032 via DeepSeek-V3 671B multi-model routing (Llama-8B for simple, V3 for standard, 70B for NSFW).
12
Plus subs to break even
6
Premium subs to break even
100+
Months runway on $500K
total burn vs $150K+/mo for a VC-backed team
v1.0 to v4.4 -- no standups, no sprint planning
in Design/UX/Psychology -- every feature is applied science
AI-native development -- Claude Code is the eng team
Founder
Lucy is built by Yudi -- a PhD + Masters in Design with deep expertise in UX, emotional design, and psychology triggers. Every feature is an applied psychology decision, not a technical afterthought.
PhD + Masters in Design, UX, Emotional Design, Psychology Triggers
v1.0 to v4.4 in 6 weeks with production-quality code
365-day intimacy arc = delayed gratification (psychology)
5-layer safety architecture, compliance-aware from day one
90% AI-native development with Claude Code
Founder & Engineer
Looking for
Seed capital to reach production launch and first 10K users. The product is built -- we need distribution.
Get In Touch
The product is built. The orchestration layer is the moat. The market is $140B and growing. The winner in AI companions will not be the best model -- it will be the best brain around the model.
Pre-seed
Stage
$140B
TAM by 2030
v4.4
Current version
No credit card required.