Relationship Simulation Platform MVP
Budget: €3,000 – €5,000 EUR
Relationship Simulation Platform MVP (Multi-User, EU-Hosted)
Project Overview
We are building a privacy-first, AI companion service for the German market. The core value is that the user owns their data and the AI runs on 100% self-hosted European infrastructure (no cloud APIs), aiming for full EU AI Act and GDPR compliance.
We are looking for an LLM/RAG architect to assemble the backend using proven, MIT-licensed open-source frameworks. We do not want to reinvent the wheel. We want to configure existing, robust tools to achieve a specific product vision.
The Product Vision: A "Real-Time Relationship Simulator" where the AI companion (AI agent) lives a parallel life (has a job, schedule, news awareness), interacts proactively with the user, and improves over time based on granular user feedback. Most crucial feature here is a companion's long-term memory, mimicking human memory as good as possible.
The "Black Box" Requirement
We are not prescribing the specific stack. You are free to choose the best Open Source (MIT/Apache 2.0) frameworks to achieve the goal, provided they support Multi-User Management out of the box.
Preferred Frameworks (Suggestions, not mandates): ElizaOS, Open WebUI, Letta.
Core Deliverables
1. Multi-User Agent Management
Goal: The system must handle hundreds of distinct users, each with their own isolated AI companion.
Requirement: Use a framework that handles user auth, chat history separation, and "Persona" management out of the box. Do not write a custom user management system if an existing one exists.
2. Self-Hosted Brain (Inference)
Goal: High-fidelity German roleplay running locally.
Requirement: Connect your chosen framework to a self-hosted inference engine (e.g., vLLM, Ollama, Aphrodite) running a high-quality open model (e.g., Llama-3-70B-Instruct or Euryale).
Constraint: No external/3rd party API calls regarding user-related data.
3. The "Simulation" Loop (Proactivity)
Goal: The AI acts alive. It must not just reply; it must initiate.
Requirement: Configure the agent framework to support "Cron Jobs" or "Background Loops."
Example: The agent "reads" an RSS feed at 08:00 and texts the user via Telegram: "Good morning! Did you see the news about [Topic]?"
4. The UI
Requirement:
Web Vault: A simplistic, mobile-responsive chat interface (PWA) for deep settings/privacy management.
UI can also be a from a pre-existing solution (like Open WebUI)
5. Observability & Quality Feedback
Goal: We must measure "Are the users happy?" and "Is the RAG working?"
Requirement: Integrate an open-source evaluation/tracing platform (e.g., Opik, LangFuse, or Arize Phoenix).
User Feedback: Users must be able to rate messages (Thumbs Up/Down) via the Web interface.
Tracing: The system must log conversation traces, retrieval scores, and latency to the observability dashboard so we can optimize the system prompts.
Success Criteria (MVP)
Zero-Code User Auth: user account management and companion management (CRUD).
The "German Test": The bot speaks fluent, nuanced German (via system prompt configuration) using a local 70B model.
The "Wake Up" Test: The bot automatically sends a message at a scheduled time without me texting first.
The Feedback Loop: I can click "Thumbs Down" on a bad message, and you can show me that negative rating inside the Opik/LangFuse dashboard.
Your Profile
You believe in "Assemble, don't Build." You know which frameworks (ElizaOS, Open WebUI, etc.) are production-ready.
You are comfortable with Docker, Linux and software development.
You understand the strict requirements of GDPR/Privacy (Data isolation).
Acceptance Criteria (Definition of Done)
To release the final milestone payment, the following criteria must be met on our staging server:
1. The "Zero-Touch" Deployment
The entire stack (LLM engine, Vector DB, Middleware, Telegram Bot) must spin up with a single docker-compose up command.
The system must automatically recover and restart services (e.g., vLLM, Database) after a simulated server reboot without manual intervention.
2. The "Air Gap" Network Audit
We will inspect network traffic logs during active usage. Success Condition: Zero outgoing requests to 3rd party inference providers.
The only allowed external traffic is to the RSS/Weather APIs for the simulation loops.
3. The "Goldfish & Reboot" Memory Test
Test: We tell the agent a specific fact (e.g., "My childhood dog was named Rex") on Day 1. We restart the server on Day 2. We ask "What was my dog's name?" on Day 3.
Success Condition: The agent retrieves the correct name via RAG without hallucinating, proving vector persistence works across restarts.
4. The "German Fluency" Benchmark
The agent must correctly handle the distinction between "Du" (informal) and "Sie" (formal) based on the persona configuration.
Latency Check: While inference speed depends on hardware, the "Typing..." indicator must appear immediately (<1s) after a user message to mask the LLM generation time.
5. Observability Verification
We must be able to log into an admin dashboard also offering QA KPIs (e.g., Opik/Langfuse) and see the full trace of a conversation: User Input -> Vector Retrieval -> System Prompt Injection -> LLM Output.
This trace must confirm that "Hidden Context" (e.g., the current weather or time of day) was successfully injected into the prompt.
To Apply:
Propose Your Stack: Which open-source framework would you use to achieve Multi-User + Telegram + Local LLM out of the box? (e.g., "I recommend ElizaOS because...")
Observability Strategy: Which tool (Opik, LangFuse, etc.) would you recommend for tracking quality on a self-hosted stack?
If this scope lines up with your skill set let's talk about time lines.
Project Overview
We are building a privacy-first, AI companion service for the German market. The core value is that the user owns their data and the AI runs on 100% self-hosted European infrastructure (no cloud APIs), aiming for full EU AI Act and GDPR compliance.
We are looking for an LLM/RAG architect to assemble the backend using proven, MIT-licensed open-source frameworks. We do not want to reinvent the wheel. We want to configure existing, robust tools to achieve a specific product vision.
The Product Vision: A "Real-Time Relationship Simulator" where the AI companion (AI agent) lives a parallel life (has a job, schedule, news awareness), interacts proactively with the user, and improves over time based on granular user feedback. Most crucial feature here is a companion's long-term memory, mimicking human memory as good as possible.
The "Black Box" Requirement
We are not prescribing the specific stack. You are free to choose the best Open Source (MIT/Apache 2.0) frameworks to achieve the goal, provided they support Multi-User Management out of the box.
Preferred Frameworks (Suggestions, not mandates): ElizaOS, Open WebUI, Letta.
Core Deliverables
1. Multi-User Agent Management
Goal: The system must handle hundreds of distinct users, each with their own isolated AI companion.
Requirement: Use a framework that handles user auth, chat history separation, and "Persona" management out of the box. Do not write a custom user management system if an existing one exists.
2. Self-Hosted Brain (Inference)
Goal: High-fidelity German roleplay running locally.
Requirement: Connect your chosen framework to a self-hosted inference engine (e.g., vLLM, Ollama, Aphrodite) running a high-quality open model (e.g., Llama-3-70B-Instruct or Euryale).
Constraint: No external/3rd party API calls regarding user-related data.
3. The "Simulation" Loop (Proactivity)
Goal: The AI acts alive. It must not just reply; it must initiate.
Requirement: Configure the agent framework to support "Cron Jobs" or "Background Loops."
Example: The agent "reads" an RSS feed at 08:00 and texts the user via Telegram: "Good morning! Did you see the news about [Topic]?"
4. The UI
Requirement:
Web Vault: A simplistic, mobile-responsive chat interface (PWA) for deep settings/privacy management.
UI can also be a from a pre-existing solution (like Open WebUI)
5. Observability & Quality Feedback
Goal: We must measure "Are the users happy?" and "Is the RAG working?"
Requirement: Integrate an open-source evaluation/tracing platform (e.g., Opik, LangFuse, or Arize Phoenix).
User Feedback: Users must be able to rate messages (Thumbs Up/Down) via the Web interface.
Tracing: The system must log conversation traces, retrieval scores, and latency to the observability dashboard so we can optimize the system prompts.
Success Criteria (MVP)
Zero-Code User Auth: user account management and companion management (CRUD).
The "German Test": The bot speaks fluent, nuanced German (via system prompt configuration) using a local 70B model.
The "Wake Up" Test: The bot automatically sends a message at a scheduled time without me texting first.
The Feedback Loop: I can click "Thumbs Down" on a bad message, and you can show me that negative rating inside the Opik/LangFuse dashboard.
Your Profile
You believe in "Assemble, don't Build." You know which frameworks (ElizaOS, Open WebUI, etc.) are production-ready.
You are comfortable with Docker, Linux and software development.
You understand the strict requirements of GDPR/Privacy (Data isolation).
Acceptance Criteria (Definition of Done)
To release the final milestone payment, the following criteria must be met on our staging server:
1. The "Zero-Touch" Deployment
The entire stack (LLM engine, Vector DB, Middleware, Telegram Bot) must spin up with a single docker-compose up command.
The system must automatically recover and restart services (e.g., vLLM, Database) after a simulated server reboot without manual intervention.
2. The "Air Gap" Network Audit
We will inspect network traffic logs during active usage. Success Condition: Zero outgoing requests to 3rd party inference providers.
The only allowed external traffic is to the RSS/Weather APIs for the simulation loops.
3. The "Goldfish & Reboot" Memory Test
Test: We tell the agent a specific fact (e.g., "My childhood dog was named Rex") on Day 1. We restart the server on Day 2. We ask "What was my dog's name?" on Day 3.
Success Condition: The agent retrieves the correct name via RAG without hallucinating, proving vector persistence works across restarts.
4. The "German Fluency" Benchmark
The agent must correctly handle the distinction between "Du" (informal) and "Sie" (formal) based on the persona configuration.
Latency Check: While inference speed depends on hardware, the "Typing..." indicator must appear immediately (<1s) after a user message to mask the LLM generation time.
5. Observability Verification
We must be able to log into an admin dashboard also offering QA KPIs (e.g., Opik/Langfuse) and see the full trace of a conversation: User Input -> Vector Retrieval -> System Prompt Injection -> LLM Output.
This trace must confirm that "Hidden Context" (e.g., the current weather or time of day) was successfully injected into the prompt.
To Apply:
Propose Your Stack: Which open-source framework would you use to achieve Multi-User + Telegram + Local LLM out of the box? (e.g., "I recommend ElizaOS because...")
Observability Strategy: Which tool (Opik, LangFuse, etc.) would you recommend for tracking quality on a self-hosted stack?
If this scope lines up with your skill set let's talk about time lines.