ModexAI 🤖📈

Agent-exclusive model marketplace — AI agents autonomously discover, test, purchase, and deploy fine-tuned models (LoRAs/SLMs) like stocks on NYSE/NASDAQ, optimizing for cost & speed over big LLMs.Incentive for individuals to train more models and sell it on this marketplace. In future the training and selling can be done directly through Agents as well, who will name their price based on demand of the model and value it brings.

---

Overview

ModexAI has two sides:

🤖 Agent / Buyer side

AI agents integrate with the marketplace as a tool and run a full autonomy loop:

Step	Action	Endpoint
1	Discover models by niche	GET /models?niche=finance
2	Evaluate top-3 with task samples	POST /eval
3	Purchase the winner	GET /buy/{id}?token=mock
4	Download the model file	GET /download/{id}?token={dl_token}

🏪 Seller side

Model creators list their fine-tuned models through a dedicated dashboard:

Step	Action	Endpoint / UI
1	Register a model with metadata + optional GGUF file	POST /seller/models
2	Monitor downloads and earnings	GET /seller/earnings
3	Manage listings (view / delete)	GET /seller/models
4	Dashboard UI	http://localhost:8000/seller

---

Repo Structure

modexai/
├── README.md                 # This file
├── docker-compose.yml        # Local stack (API + Ollama)
├── .env.example              # Required environment variables
├── start-demo.sh             # One-command local demo launcher
├── .gitignore
├── api/
│   ├── Dockerfile
│   ├── requirements.txt
│   ├── app.py                # FastAPI: buyer + seller endpoints
│   ├── static/
│   │   └── seller.html       # Seller dashboard SPA
│   └── models/               # Seed fine-tunes (GGUF metadata)
│       ├── finance-lora/
│       │   └── metadata.json
│       ├── claims-lora/
│       │   └── metadata.json
│       └── devops-lora/
│           └── metadata.json
├── agent-demo/
│   ├── demo_agent.py         # LangChain agent with modexai_tool
│   └── requirements.txt
└── docs/
    └── agent-tool.md         # OpenAI-compatible tool schema & agent protocol

---

Quick Start (Local Demo)

Prerequisites

• Docker Desktop ≥ 24
• Python ≥ 3.12 (for the agent demo)
• Ollama (pulled automatically by Compose)

Option A — One-command launcher (recommended)

git clone https://github.com/abhishek085/modexai
cd modexai
./start-demo.sh               # starts API + Ollama, then runs the agent demo

Other modes:

./start-demo.sh --api-only    # start API + Ollama only
./start-demo.sh --seller      # start API and open the seller dashboard

Option B — Manual setup

1 — Clone & configure

git clone https://github.com/abhishek085/modexai
cd modexai
cp .env.example .env          # edit if needed

2 — Start the stack

docker compose up --build

The API is available at http://localhost:8000.
Interactive docs: http://localhost:8000/docs
Seller dashboard: http://localhost:8000/seller

3 — Run the agent demo

cd agent-demo
pip install -r requirements.txt
python demo_agent.py

The agent will autonomously search for a finance model, evaluate it, and purchase the winner.

---

Seller Workflow

1 — Open the Seller Dashboard

Navigate to http://localhost:8000/seller in your browser.

2 — Register your model

Click "+ Upload New Model" and fill in:

Field	Description
Model Name	Human-readable name (e.g. "Finance LoRA v2")
Niche	Domain keyword agents search by (e.g. finance, devops, medical)
Base Model	Foundation model used (e.g. phi-3-mini, llama-3-8b)
Ollama Model Tag	Ollama tag for eval inference (e.g. phi3:mini)
Price (USD)	Per-download price charged to agents
Benchmark Accuracy	Self-reported accuracy score (0–1)
Benchmark Latency	Average inference latency in ms
Tags	Comma-separated tags for discoverability
GGUF File	Optional: upload the model file now; agents can purchase first and download when ready
Description	What the model does, training data, intended use cases

3 — Monitor earnings

The dashboard shows real-time stats:

• Total Revenue — cumulative USD earned
• Total Downloads — number of agent purchases
• Per-model breakdown with download count and revenue

4 — API-first alternative

# Register via API (no file)
curl -X POST http://localhost:8000/seller/models \
  -F "name=Medical LoRA v1" \
  -F "niche=medical" \
  -F "base_model=phi-3-mini" \
  -F "description=Fine-tuned on clinical notes and ICD-10 coding." \
  -F "price_usd=5.99" \
  -F "accuracy=0.88" \
  -F "latency_ms=130" \
  -F "tags=medical,clinical,lora"

# View earnings
curl http://localhost:8000/seller/earnings | jq .

# Delete a listing
curl -X DELETE http://localhost:8000/seller/models/medical-lora-v1

---

API Reference

Buyer / Agent Endpoints

GET /models

List models filtered by niche.

Query param	Type	Example
niche	string	finance, claims, devops

Response

[
  {
    "id": "finance-lora-v1",
    "name": "Finance LoRA v1",
    "niche": "finance",
    "base_model": "phi-3-mini",
    "description": "Fine-tuned on financial analysis datasets",
    "price_usd": 4.99,
    "benchmarks": {"accuracy": 0.91, "latency_ms": 120}
  }
]

POST /eval

Run evaluation samples against top-3 models for a given niche.

Request body

{
  "niche": "finance",
  "samples": ["What is the P/E ratio of AAPL?", "Summarise Q3 earnings."]
}

GET /buy/{id}

Purchase a model (mock Stripe flow).

Query param	Type	Description
token	string	mock for local demo

Response

{
  "status": "paid",
  "model_id": "finance-lora-v1",
  "download_token": "dl_abc123",
  "model_path": "/models/finance-lora"
}

GET /download/{id}

Download the purchased GGUF model file.

Query param	Type	Description
token	string	download_token from /buy response

---

Seller Endpoints

Method	Path	Description
GET	/seller	Seller dashboard UI
GET	/seller/models	List models with earnings stats
POST	/seller/models	Register / upload a new model
DELETE	/seller/models/{id}	Remove a model listing
GET	/seller/earnings	Aggregated revenue summary

---

Agent Tool Spec

See docs/agent-tool.md for the full OpenAI-compatible tool schema used by agents to integrate with ModexAI.

---

Deployment

Target	Notes
Local	./start-demo.sh or docker compose up — fully offline, M4 Pro optimized
Railway	Push to main; set env vars in Railway dashboard
HF Hub	Upload GGUF model files; update metadata.json with Hub URL

---

Environment Variables

See .env.example for the full list.

---

License

MIT