Provides streaming chat capabilities using OpenAI's GPT models with automatic retry logic and rate limit handling for conversational AI tasks.
๐ง MCP Server (Model Compute Paradigm)
A modular, production-ready FastAPI server built to route and orchestrate multiple AI/LLM-powered models behind a unified, scalable interface. It supports streaming chat, LLM-based routing, and multi-model pipelines (like analyze โ summarize โ recommend) โ all asynchronously and fully Dockerized.
๐ฏ Project Score (Production Readiness)
Capability | Status | Details |
๐ง Multi-Model Orchestration | โ Complete | Dynamic routing between
,
,
,
|
๐ค LLM-Based Task Router | โ Complete | GPT-powered routing via
task type |
๐ Async FastAPI + Concurrency | โ Complete | Async/await + concurrent task execution with simulated/model API delays |
๐ GPT Streaming Support | โ Complete |
chunked responses for chat endpoints |
๐งช Unit + Mocked API Tests | โ Complete | Pytest-based test suite with mocked
responses |
๐ณ Dockerized + Clean Layout | โ Complete | Python 3.13 base image, no Conda dependency, production-ready Dockerfile |
๐ฆ Metadata-Driven Registry | โ Complete | Model metadata loaded from external YAML config |
๐ Rate Limiting & Retry | โณ In Progress | Handles 429 retry loop; rate limiting controls WIP |
๐งช CI + Docs | โณ Next | GitHub Actions + Swagger/Redoc planned |
๐งฉ Why This Project? (Motivation)
Modern ML/LLM deployments often involve:
Multiple task types and model backends (OpenAI, HF, local, REST)
Routing decisions based on input intent
Combining outputs of multiple models (e.g.,
summarize
+recommend
)Handling 429 retries, async concurrency, streaming responses
๐ง However, building such an LLM backend API server that is:
Async + concurrent
Streamable
Pluggable (via metadata)
Testable
Dockerized โฆ is non-trivial and not easily found in one single place.
๐ก What Weโve Built (Solution)
This repo is a production-ready PoC of an MCP (Model-Compute Paradigm) architecture:
โ FastAPI-based microserver to handle multiple tasks via
/task
endpointโ Task router that can:
๐ Dispatch to specific model types (
chat
,sentiment
,summarize
,recommend
)๐ค Use an LLM to infer which task to run (
auto
)๐ง Run multiple models in sequence (
analyze
)
โ GPT streaming via
text/event-stream
โ Async/await enabled architecture for concurrency
โ Clean modular code for easy extension
โ Dockerized for deployment
โ Tested using Pytest with mocking
๐ ๏ธ Use Cases
Use Case | MCP Server Support |
Build your own ChatGPT-style API | โ
task with streaming |
Build intelligent task router | โ
task with GPT-powered intent parsing |
Build AI pipelines (like RAG/RL) | โ
task with sequential execution |
Swap between OpenAI/HuggingFace APIs | โ Via
config |
Add custom models (e.g., OCR, vision) | โ Just add a new module + registry entry |
๐ Features
โ Async FastAPI server
๐ง Task-based Model Routing (
chat
,sentiment
,recommender
,summarize
)๐ Model Registry from YAML/JSON
๐ Automatic Retry and Rate Limit Handling for APIs
๐ Streaming Responses for Chat
๐งช Unit Tests + Mocked API Calls
๐ณ Dockerized for production deployment
๐ฆ Modular structure, ready for CI/CD
๐ Architecture Overview
This server cannot be installed
remote-capable server
The server can be hosted and run remotely because it primarily relies on remote services or has no dependency on the local environment.
Enables routing of ML tasks like chat, sentiment analysis, recommendations, and summarization to appropriate models through a dynamic YAML-based registry. Provides async FastAPI endpoints with streaming support, retry logic, and pluggable model architecture for scalable ML inference.