Building an AI Companion Like Rosely AI with ChatGPT

Discussions
,
1

Hi everyone,

I’ve been working on an AI companion platform inspired by apps like Rosely AI. My goal isn’t to clone the product feature-for-feature, but to understand how modern AI companion applications are architected and how to build something scalable using current AI APIs.

Instead of training my own large language model, I decided to use the ChatGPT API as the conversational engine and focus on building the surrounding infrastructure.

I’d love to get some feedback from developers who have built AI-powered applications or worked with LLMs in production.

Current Tech Stack

Frontend

  • Next.js
  • TypeScript
  • Tailwind CSS

Backend

  • Node.js
  • Express
  • PostgreSQL
  • Redis

AI Services

  • ChatGPT API
  • AI Image Generation API
  • Text-to-Speech API

Features Implemented

The application currently includes:

  • AI chat with multiple personalities
  • Conversation history
  • User authentication
  • AI-generated images
  • Voice replies
  • Subscription system
  • Mobile-friendly interface

One thing I noticed while studying platforms like Rosely AI is that the overall experience depends much more on memory and personalization than simply using a powerful language model.

Memory Architecture

This has probably been the most interesting part of the project.

Rather than sending the complete chat history with every request, I’m testing a layered memory system.

Current flow:

  • Recent conversation for context
  • User profile and preferences
  • Long-term memory summaries
  • Retrieved memories relevant to the current message

This significantly reduces token usage while making conversations feel more consistent.

Has anyone here tried retrieval-based memory or another approach?

Prompt Engineering

Another challenge has been prompt management.

Currently each request is built from:

  • System instructions
  • Character personality
  • User profile
  • Memory context
  • Recent conversation
  • Current message

Even small prompt adjustments noticeably change the AI’s behavior.

I’d be interested to hear how others structure prompts for long-running conversations.

Performance & Scaling

As I add more concurrent users, API latency becomes more noticeable.

I’m currently exploring:

  • Streaming responses
  • Redis caching
  • Background job queues
  • Horizontal scaling
  • Better database indexing

If you’ve optimized an AI application, what had the biggest impact on response time?

Image Generation

Image generation runs as a separate service rather than inside the chat pipeline.

The workflow is:

  1. User requests an image.
  2. Backend validates the request.
  3. AI image service processes it.
  4. Image is stored and returned asynchronously.

This keeps the chat interface responsive while images are being generated.

Looking for Advice

I’d appreciate feedback on:

  • AI memory systems
  • Prompt engineering
  • API cost optimization
  • Scaling Node.js applications
  • WebSocket architecture
  • AI moderation
  • Database design for AI chats

If you’ve built an AI chatbot, AI companion, or another LLM-powered SaaS application, I’d love to hear what worked well and what challenges you faced.

1 Like