Skip to main content

Getting Started

Protean AI is an end-to-end enterprise AI platform designed to help teams fine-tune, deploy, and operate domain-specific models in a sovereign, cost-efficient, and secure way. This section introduces the core concepts and the typical workflow, enabling a structured path from raw data to deployed AI agents.

What Protean AI Provides​

Protean AI brings together all essential building blocks required for enterprise AI development:

  • Node Registration & Management: Secure registration and management of compute nodes (CPU/GPU) that are authorized to run training jobs and serve inference workloads. Nodes are the trusted compute boundary of the platform, so registering them is the prerequisite that lets the system place training and inference work onto known, approved hardware rather than arbitrary machines.
  • Model Registry: Centralized management of base models, quantized variants, versions, and metadata. It acts as the single source of truth for which models exist and how they relate, so every fine-tuning run and deployment starts from a known, versioned artifact instead of an ad hoc copy.
  • Fine-Tuning Engine: Opinionated and efficient training pipelines with strict alignment between dataset format, objective, loss function, and evaluation strategy. Because these pieces are kept consistent with one another, training avoids the silent misconfiguration that produces misleading results, which makes runs reproducible and comparable.
  • Adapter Registry: A dedicated registry for managing adapters (LoRA/QLoRA) independent of base models. It supports uploading adapters in standard HuggingFace format or publishing fine-tuned checkpoints as reusable adapters, and it tracks versioning, metadata, and lineage. Keeping adapters separate from base models lets the same specialization be composed with different models at deployment time, which avoids retraining full weights for each variant.
  • Model Runtime: Lightweight, scalable inference with support for adapters, quantization, and streaming responses. The runtime serves deployed models on registered nodes, and its support for quantization and adapters keeps the memory footprint and infrastructure cost low while still allowing specialized behavior.
  • Knowledge set Management: Native support for preparing, versioning, and validating Retrieval-Augmented Generation datasets, including document ingestion, chunking strategies, and embedding generation. Managing these steps as a versioned, validated set makes retrieval behavior predictable, so changes to source documents or chunking do not silently degrade answer quality.
  • RAG APIs: Configurable retrieval pipelines that combine vector search, ranking, and prompt assembly in a reproducible way. They let retrieval be defined and validated independently of generation, which means retrieval quality can be tuned and verified before it ever influences a model's output.
  • MCP Server: Support for both internal and external Model Context Protocol (MCP) servers. External MCP servers can be registered with managed authorization and access control, while Protean AI provides a built-in SQL MCP server for secure, structured access to relational data. This gives agents a governed way to reach tools and data sources, so structured access stays under access control rather than being wired in ad hoc.
  • Agent Framework: Tools for building, deploying, and operating agents that combine models, RAG pipelines, MCP server integration, tool execution, and memory management. It is the integration layer that ties the other building blocks together, so a deployed agent can reason over retrieved context and act through tools within controlled boundaries.
  • Governance & Control: Built-in support for data sovereignty, security, auditability, and operational safety. These controls run across every resource, so enterprise requirements around where data lives, who can access it, and what is auditable are enforced by the platform rather than bolted on afterward.

Workflow​

A standard Protean AI workflow follows a predictable and reproducible sequence:

  1. Register a Node
    Register a compute node with Protean AI before any deployment or training can occur.
    Node registration establishes trust, advertises available hardware (CPU, GPU, memory), and defines where models are allowed to run.

  2. Select a Base Model
    Choose a foundational or instruction-tuned model that matches the task, domain complexity, and registered hardware constraints.

  3. Prepare Training Datasets
    Structure supervised training data using supported dataset formats aligned with the training objective (chat, classification, embedding, reranking, etc.).

  4. Configure and Run Fine-Tuning
    Define objective, evaluation metrics, and efficiency settings such as LoRA/QLoRA, gradient accumulation, etc.
    Execute training with frequent evaluation and reproducible results.

  5. Publish the Adapter
    Publish the fine-tuned checkpoint as an adapter for reuse across multiple base models deployments. Upload an existing adapter in HuggingFace format into the Adapter Registry. Adapters are managed independently of base models, enabling flexible composition and reuse. Adapter Registry manages versioning, lineage, and metadata of adapters.

  6. Prepare Knowledge Sets Ingest enterprise documents and structure them into Knowledge Sets. Apply deterministic chunking strategies, embedding models, and metadata enrichment to ensure predictable retrieval behavior.

  7. Configure RAG
    Define retrieval, ranking, and prompt-assembly behavior using the RAG APIs.
    Validate retrieval quality independent of generation.

  8. Deploy Models for Inference
    Deploy models and adapters to registered nodes using the runtime, with optional quantization to minimize memory footprint and infrastructure cost.

  9. Build and Deploy Agents
    Compose agents that combine deployed models, RAG pipelines, and tools.
    Deploy agents as long-running services that can reason, retrieve context, and act within controlled boundaries.

Note: Access to all the resources and actions are strictly controlled by the authorization policies defined at the individual resource level.

Hardware and Efficiency First​

Protean AI is designed for lean and sovereign deployments:

  • Quantization support to significantly reduce VRAM and hardware requirements.
  • Adapter-based fine-tuning to avoid retraining full model weights.
  • Efficient runtimes suitable for on-prem, private cloud, VPC, and edge environments.

This enables expert models to run close to enterprise data without risks of data exfiltration.

Developer Experience​

Protean AI integrates naturally into existing software development workflows:

  • Explicit configuration with clear intent.
  • Constrained dataset formats that prevent training-time ambiguity.
  • Reproducible experiments and comparable results across runs.
  • Production-oriented APIs for models, RAG, and agents.

Next Steps​

To continue:

  • Register a Node and verify available hardware.
  • Explore the Model Registry for base models and checkpoints.
  • Upload or publish an Adapter for efficient specialization.
  • Review Training Dataset Formats to align data with objectives.
  • Follow the Fine-Tuning guide to train a first model or adapter.
  • Prepare a Knowledge set and configure retrieval pipelines.
  • Build and deploy an Agent as the final integration step.

This foundation ensures every step, from raw data and infrastructure to deployed agents, is intentional, efficient, and enterprise-ready.