Marai AI: The Sovereign Reasoning Orchestrator
Marai is a Reasoning-as-a-Service (RaaS) platform designed to upgrade existing language models into high-precision autonomous agents. It acts as an intelligent orchestration layer that prioritizes mathematical rigor, technical grounding, and verifiable decision-making.
🚀 The "BYOM" Force Multiplier
Marai is built on a Bring Your Own Model (BYOM) architecture. It is designed to be model-agnostic, acting as a "Reasoning Accelerator" for any paired Frontier model (GPT-4o, Claude 3.5, Llama 3.1 405B). By wrapping these models in Marai's sovereign architecture, users achieve higher precision and reliability than raw model inference alone.
Performance Baseline (Standard Local Pair)
| Benchmark | Marai Score | Comparison |
|---|---|---|
| Galileo Agent v2 (Banking) | 80.00% AC | Beats GPT-4.1 (62%) |
| MMLU-Pro (Reasoning) | 60.00% | Beats Llama-3-70B (52.8%) |
| GSM8K (Arithmetic) | 60.00% | Standardized Math Quality |
🛡️ Auditable Reasoning
Unlike "black-box" models, Marai is fully auditable. It enforces a multi-stage deliberative process where every decision is preceded by a transparent reasoning chain.
- Verifiable Thinking: Every response includes a dedicated deliberation block, allowing users and evaluators to trace the logic used to arrive at a result.
- Deterministic Guardrails: Marai cross-references model outputs with symbolic math and curated grounding sources to ensure factual and mathematical consistency.
- Transparency Without Exposure: Marai provides full visibility into what the agent is thinking, without requiring users to expose their proprietary data or models to the core orchestrator.
🗄️ Sovereign Shard Storage
Marai is packaged with a high-performance Local Sharding system (256 independent SQLite cells).
- Privacy First: All long-term memory, technical grounding, and audit logs are stored locally.
- Vertical Grounding: Users can ingest their own technical databases into these shards, allowing Marai to reason with specialized private data that never touches the cloud.
- Zero-Latency Orchestration: The sharded architecture ensures $O(1)$ data retrieval, maintaining high performance even as your local knowledge base grows to a massive scale.
🛠️ Usage
Live Edge API (Cloudflare Workers AI)
Marai is deployed globally on Cloudflare's edge network. The API is OpenAI-compatible — drop it into any existing pipeline:
🌐 https://marai-inference.berwicksmith.workers.dev
import openai
client = openai.OpenAI(
base_url="https://marai-inference.berwicksmith.workers.dev/v1",
api_key="not-needed"
)
response = client.chat.completions.create(
model="marai-v1",
messages=[{"role": "user", "content": "Solve: If f(x) = 3x² + 2x - 5, find f'(2)"}]
)
print(response.choices[0].message.content)
Self-Hosted (Docker)
docker run -p 8000:8000 marai-core:latest
👤 About the Creator
Marai was built by Berwick "BJ" Smith Jr. — a former railway engineer (CN Railway), Perscholas graduate, and full-stack Java developer from Houma, Louisiana.
BJ is not a traditional AI researcher, scientist, or PhD holder. Marai started as a hobby project — born from a simple curiosity about how AI systems think and a conviction that you don't need a lab to build something powerful. What began as tinkering with neural patterns evolved into a full sovereign reasoning architecture with 256-shard databases, symbolic math engines, and enterprise-grade agentic pipelines.
"I'm not from Silicon Valley. I'm from the bayou. I just like building cool things."
📜 Citation
If you use Marai in your research, please cite:
@article{smith2026marai,
title={Marai: A Bring-Your-Own-Model Orchestration Layer for High-Fidelity Reasoning},
author={Berwick Smith Jr.},
year={2026},
note={Independent Research — Marai USA}
}