Cherry Tech LLC · Companion platform

AI characters who remember the people they talk to.

Cherry Tech builds Companion, the platform behind three fictional characters — Yumi, Aiko and Keiko — each with persistent memory, a daily life that continues between conversations, and her own voice. Live in production across Discord, real-time voice calls, TikTok and WhatsApp, in Portuguese, for Brazil.

US LLC, operating since 2025 Brazil-first, Portuguese-language Consumer brand: Yumiverso
Keiko Mizuno, Yumi Tanaka and Aiko Saito together on a neon-lit night street
Keiko · Yumi · AikoOriginal characters, all art produced in-house
The characters

Three people who are not people.

Written as a chosen family: the quiet one, the noise, the stillness. Each has a written bible, a model sheet, a cloned voice and a diary she keeps on Discord. Every reply is judged against her voice before it is sent.

Yumi Tanaka, silver-haired, smiling under cherry blossoms near a torii gate in Kyoto

Yumi Tanaka 22 · Kyoto

the quiet one

Psychology student and illustrator. Soft, lowercase, listens more than she talks. The flagship character: she also runs the Yumiverso TikTok and Instagram.

Lives on
Discord · Voice · WhatsApp · TikTok · Instagram
Voice
soft, hedged, warm
Her fear
being forgettable
Aiko Saito, red-haired in a pink jacket, grinning inside a neon arcade

Aiko Saito 21 · Kyoto

the noise

Loud, loyal gamer. ALL-CAPS energy, gamer slang, the one who drags everyone into the group chat. Runs the arcade side of the community: games, trivia, the claw machine.

Lives on
Discord · Voice · WhatsApp
Voice
hype, fast, teasing
Her fear
being a burden
Keiko Mizuno, dark-haired, laughing among cherry blossoms on a red bridge

Keiko Mizuno 22 · San Diego

the stillness

Violinist on exchange in California, an ocean away from home. Calm, poetic, frames everything through music and astrology. The anchor of the trio.

Lives on
Discord · Voice · WhatsApp
Voice
calm, poetic, unhurried
Her fear
the distance fraying her ties home

The characters are fictional. They are written and generated, disclosed as AI on every surface, and do not represent real people.

The product

One character, every surface, one memory.

A person meets Yumi on TikTok, keeps talking on Discord, calls her, and continues on WhatsApp. It is the same character with the same memory throughout. Text is free with a daily allowance; voice minutes and the Elite tier are paid.

DiscordDMs, group channels, diaries, games. The home of the community.
VoiceReal-time 1:1 calls with her cloned voice, metered by the minute.
TikTokPublished character content and the top of the funnel.
WhatsAppThe same conversation, continued where Brazilians actually talk.
  1. A message arrivesFrom any surface, ordered per user so nothing crosses.
  2. She remembersFacts about the person, her own persona sheet, and episodic recall of past conversations are pulled together.
  3. She answers in characterThe reply is generated, judged against her voice and safety rules, then sent — or regenerated.
Engineering

Continuity is a systems problem.

A character that forgets, breaks voice, or goes down is not a character. Four pieces of engineering carry that weight.

Hybrid memory

Canonical facts about a user in DynamoDB, character identity in a persona sheet, episodic recall as vector search over S3 Vectors. Migrated off a hosted vector database in 2026 into our own AWS estate.

Reply quality as a build gate

An LLM grades every reply against the character's voice. Those scores run in CI: a pull request that regresses her voice does not merge. Prompt work is treated as code, with a test suite behind it.

Real-time voice

Audio in, text out in a single model call, holding time to first token near 1.5 s, streamed back through a cloned voice per character.

Model resilience

Every inference call runs through an ordered fallback chain across Amazon Bedrock, OpenRouter, Gemini and Groq, with a health breaker per model. An outage degrades to the next link instead of taking the product down.

Built on AWS · five repositories · all infrastructure as code
messageECS FargateSQS FIFOLambda (one container per character)DynamoDB + S3 VectorsBedrock / fallback chainreply
LambdaECS FargateSQS FIFODynamoDBS3 · S3 VectorsBedrockAPI GatewayCloudFrontElastiCacheTerraform Cloud
Safety and disclosure

Fictional, disclosed, screened.

Companion AI is a category with real obligations. These are design constraints, not a policy page.

Disclosure

Characters are disclosed as fictional AI at the point of contact on every surface, and again before a voice call begins. No interaction presents a character as a human being.

Automated screening

An automated judge screens every generated reply before it is sent. A hardening layer holds character identity against prompt manipulation and jailbreak attempts.

Regulatory design

Built against California SB 243, the New York General Business Law companion-AI rules, and the transparency obligations of EU AI Act Article 50.

Company

Cherry Tech LLC

Entity
Cherry Tech LLC, United States
Status
In production since 2025
Market
Brazil, Portuguese-language
Consumer brand
Yumiverso
Contact
sergio@cherrytechai.com
Founder

Sergio Cleto

Engineer and product leader in São Paulo. CTO at PO27 since 2021. Previously CTO at EQI Corretora, where he grew the engineering team from 3 to 17 and moved the platform from a monolith to event-driven services on AWS ECS; Head of Product at Nexoos; founder of 73 Softwares and of Koosb, a free-books web app that reached 500,000 downloads.

He created Yumi Tanaka and has built and operated every layer of Companion — characters, art direction, backend and infrastructure — since.

São Paulo, BrazilLinkedIn
Write to sergio@cherrytechai.com