Distributed AI Ensemble Public Research Portal

Start Here

Use this public portal in three simple steps.

1

Read the Architecture

Open the interactive architecture map to understand query flow from orchestrator to agents.

2

Try a Live Query

Go to Playground, set your public orchestrator URL, and run one strategy with fixed seed.

3

Check Results and Paper

Review benchmark charts, CSV/JSON exports, and the full IEEE manuscript from this same site.

Abstract

This project presents a complete local distributed AI network in which one FastAPI orchestrator routes prompts to four local Ollama-based LLM agents, aggregates answers with multiple ensemble strategies, and logs benchmark-ready metrics. The system implements majority voting, dynamic weighted voting, inverse surprising popularity (ISP), topic-based routing, and two-round debate. Evaluation is automated over MMLU, GSM8K, and TruthfulQA with deterministic controls, repeated trials, confidence intervals, paired t-tests, and Wilcoxon signed-rank tests. All outputs are exported as CSV, JSON, PNG figures, and IEEE LaTeX tables, with an accompanying manuscript package for conference submission.

Research Video Walkthrough

Watch the full visual explanation of this distributed AI ensemble project, including architecture flow, live multi-agent querying, optimization, and benchmark outcomes.

Why watch this video?

Quickly understand the full end-to-end system flow.
See how all agent models collaborate in real time.
Review benchmark outputs and statistical reporting.

Open Video on YouTube

System Architecture

Node Layout

Orchestrator: 172.16.185.223
Agent 1: 172.16.185.209 (llama3.2:3b)
Agent 2: 172.16.185.218 (qwen2.5:3b)
Agent 3: 172.16.185.220 (phi3:mini)
Agent 4: 172.16.185.222 (gemma2:2b)

Aggregation Methods

Majority Voting
Dynamic Weighted Voting
Inverse Surprising Popularity (ISP)
Topic-Based Routing
Two-Round Debate

Click any node in the architecture map to view role details.

Live AI Playground

This playground calls your orchestrator /query API directly. For public use, set a public HTTPS endpoint (for example, your tunnel URL) with CORS enabled on the orchestrator.

Easy Steps

For this deployment, use https://ai.marcbd.site as the Orchestrator API URL.
Click Check Connection to test /health.
Select one or more models in Select Models (you can compare models individually).
Recommended stable test settings: Strategy majority, Seed 42, Temperature 0.0, Deterministic ON, Max tokens 24-64.
Click Run Query and review the final answer, JSON output, and model charts below.

Support Endpoint

Recommended public endpoint for this project:

https://ai.marcbd.site

Open /health

Orchestrator API URL

Example endpoint: https://your-subdomain.example.com (without trailing /query).

Select Models (individual control)

Click Check Connection to load models from /agents.

Strategy Temperature Max tokens Seed Deterministic mode Prompt

Response

Waiting for query...

No response yet.

Live Query Insights

Final Answer -

Agreement -

Total Latency -

Active Models -

Agreement distribution will appear after query.

Per-Model Latency (ms)

Per-Model Token Count

Answer Distribution by Model

Per-Model Result Table

Agent	Model	Answer	Latency (ms)	Tokens	Status
Run a query to populate per-model results.

Paper Access

Only the public abstract is shown by default. Full manuscript reading and downloads are restricted behind your follow, subscribe, permission-request, and password gate.

Public Abstract

This project presents a complete local distributed AI network in which one FastAPI orchestrator routes prompts to four local Ollama-based LLM agents, aggregates answers with multiple ensemble strategies, and logs benchmark-ready metrics. The system implements majority voting, dynamic weighted voting, inverse surprising popularity, topic-based routing, and two-round debate. Evaluation is automated over MMLU, GSM8K, and TruthfulQA with deterministic controls, repeated trials, confidence intervals, paired t-tests, and Wilcoxon signed-rank tests. All outputs are exported as CSV, JSON, PNG figures, and IEEE LaTeX tables, with an accompanying manuscript package for conference submission.

Restricted Access Policy

Visitors must follow my GitHub, subscribe to my YouTube channel, send a permission request, and then enter the approved access password to download the protected paper package.