Platform Overview

Overview of MAXSO AI enterprise infrastructure software matrix, system architecture, and private deployment solutions.

MAXSO AI delivers production-ready large language model infrastructure software systems for modern enterprises. We focus on solving the core engineering challenges when adopting large language models (LLMs):

Engineering complexity of private model customization and fine-tuning
Security perimeters and compliance governance barriers

Why Choose MAXSO AI

Enterprises often face uncontrollable security risks, fragmented engineering toolchains, and brittle open-source integrations when adopting large models. MAXSO AI provides key advantages:

100% Private & Air-gapped Ready: Supports complete offline deployment with zero external telemetry, strictly adhering to Zero Trust security architecture so proprietary data never leaves the premises.
Out-of-the-Box, 80% Lower Barrier: Replaces obscure distributed command-line scripts and custom proxy code with intuitive, workflow-driven enterprise software.
Standard Protocol Compatibility & Transparent Integration: The Security Gateway natively supports the OpenAI API specification. Existing systems and open-source frameworks (such as LangChain) can integrate within seconds by simply redirecting the endpoint.

Enterprise Product Matrix

MAXSO AI focuses on two core enterprise software systems:

Product	Core Value & Positioning
LLM Visual Training	A lightweight AI model fine-tuning tool for non-professional users. Through a minimalist three-step process (preparation → training → testing), even novice users can complete the entire process from data preparation to model validation.
LLM Security Gateway	Security proxy between enterprise applications and LLMs, providing prompt firewalls, bidirectional PII masking, token quota governance, and jailbreak interception

Typical Enterprise Scenarios

Scenario 1: Domain-Specific Model Fine-Tuning & Knowledge Retention

Challenge: Generic foundation models lack domain terminology and proprietary business context, while algorithm engineering fine-tuning tools present high hurdles.
Solution: Use MAXSO Visual Training to upload QA pairs, launch training with one click, monitor loss convergence curves, and perform double-blind evaluations.

Scenario 2: Preventing Data Leakage When Accessing Public Models

Challenge: Employees interacting with external models may unintentionally paste customer information, credentials, or confidential business data.
Solution: Deploy the MAXSO Security Gateway across the enterprise so all outbound requests automatically detect and mask PII with secure placeholders, preventing data exfiltration.

Scenario 3: Compliance Proxy & Quota Control for Internal LLMs

Challenge: Internal vLLM or Ollama deployments lack department-level access control, rate limiting, token cost accounting, and jailbreak audit trails.
Solution: Use the MAXSO Security Gateway as the unified internal entrance to distribute API keys, enforce rate limits, track token usage, and record audit logs.

Documentation Roadmap

Explore the Quickstart Guide for onboarding paths.
View LLM Visual Training for dataset specifications and training guides.
View LLM Security Gateway for defense policies and security rules.

Quickstart Guide

Quick introduction to MAXSO AI core products with direct links to product pages, live demos, and technical documentation.

MAXSO AI provides private large language model infrastructure software for enterprise scenarios. Depending on your operational requirements, jump directly to the relevant product documentation to learn more, or visit the product overview pages to explore features and try the products directly.

LLM Visual Training

Documentation: Go to LLM Visual Training Documentation to view product documentation
Product Overview & Live Demo: Visit LLM Visual Training Product Page to explore features, interface previews, and Try Now

LLM Security Gateway

Documentation: Go to LLM Security Gateway Documentation to view product documentation
Product Overview & Live Demo: Visit LLM Security Gateway Product Page to explore protection architecture, deployment models, and Try Now

Further Assistance

For deployment planning, hardware sizing, or enterprise integration questions, visit Support & FAQs or contact our engineering support team.

LLM Visual Training (ProtoFlow)

A one-stop no-code large model fine-tuning platform for enterprise engineers and business experts.

1. Product Overview and Positioning

1.1 Product Positioning

Enterprises rarely lack ideas for AI applications. What they need is a fast, low-cost way to answer a more practical question: Can our business data be used to train a model that is truly useful? ProtoFlow is a visual AI model training platform designed for enterprise teams. It consolidates data preparation, LoRA fine-tuning, and performance comparison into an intuitive three-step browser workflow (Prepare → Train → Test), enabling business leaders, AI leads, and engineering teams to validate concepts and drive projects forward with minimal trial-and-error overhead.

Core Positioning: Create through prototypes — complete an understandable, comparable, and reviewable fine-tuning validation loop in three clear steps: Prepare → Train → Test.

1.2 Core Value Proposition

Start with existing datasets or corporate documents, configure models with recommended templates, track training convergence visually, and evaluate base versus fine-tuned models side by side within a single conversational workspace.

Key Questions	How ProtoFlow Solves It
Is this business scenario worth fine-tuning with AI?	Validate hypotheses quickly using real business data through reviewable training runs and chat testing.
How can non-algorithm teams actively participate?	Configure, train, and test via a visual three-step workflow without writing complex code.
How do we verify whether fine-tuning was effective?	Compare base and fine-tuned models side by side with real-time training metrics for quantitative assessment.
Can proprietary data remain strictly on-premises?	Supports server-side or local deployment; all data, training, and inference can stay inside the intranet.

1.3 Traditional Fine-Tuning vs. ProtoFlow

Traditional fine-tuning forces teams to navigate Python dependencies, CUDA drivers, distributed training frameworks, hyperparameter tuning, GPU cluster provisioning, and bespoke evaluation scripts. ProtoFlow packages these capabilities into a cohesive graphical platform:

Stage	Traditional Approach	ProtoFlow
Preparation	Build environments, write training scripts, configure PyTorch	Select data, models, and training presets directly in the browser
Training Process	Troubleshoot status through raw command-line logs	Inspect live task progress, Loss convergence curves, and gradient norms
Result Validation	Author custom inference or evaluation scripts	Test interactively in chat and compare answers side by side
Dataset Construction	Manually scrape, clean, and reformat records	Upload existing data or automatically generate Q&A datasets from docs

2. Core Capabilities

2.1 Data Preparation and Intelligent Processing

Import existing JSON/JSONL datasets directly or ingest common documents (PDF, DOCX, Markdown, Excel, PPT) into the document library. The platform automatically extracts structure, segments content semantically, and leverages LLMs to generate high-quality question-and-answer (Q&A) datasets, drastically reducing manual annotation overhead.

Dataset Management & Upload: Supports selecting existing datasets and uploading JSON/JSONL files up to 100MB per file.
Document-to-Dataset Generation: Automated document parsing, semantic chunking, batch Q&A pair extraction, and data synthesis.
Data Quality Management: Duplicate detection, length filtering, format normalization, LLM-driven quality scoring, dataset merging, and audit records.

2.2 Model Configuration and Lightweight Fine-Tuning

Select pre-integrated foundation models and apply scenario templates with recommended hyperparameters to initiate training quickly. The platform natively supports leading open-source model families (such as Qwen3 and DeepSeek-R1) and offers intuitive controls for key hyperparameters.

Pre-Integrated Models & Templates: Built-in support for Qwen3, DeepSeek-R1, and other popular open-source architectures, along with templates for customer support, enterprise Q&A, and document summarization.
Flexible Hyperparameter Tuning: Fine-tune learning rate, training epochs, batch size, cutoff length, LoRA rank, weight decay, and warmup ratio.
Training Monitoring & Lifecycle Management: Real-time display of execution state, loss convergence curves, and gradient norms; supports checkpoint saving, training suspension, and one-click model weights export.

2.3 Evaluation and Side-by-Side Comparison

Launch an inference instance with one click upon training completion and conduct multi-turn comparative dialogues between the base foundation model and the fine-tuned model within a unified chat view.

Side-by-Side Comparison: Dual-column layout displays streaming, multi-turn dialogues from both base and fine-tuned models simultaneously.
Dynamic Sampling Tuning: Adjust sampling parameters such as Temperature, TopP, TopK, and MinP on the fly; inspect chain-of-thought reasoning when supported by the model.
Multidimensional Assessment: Visually evaluate domain knowledge absorption, tone consistency, and instruction compliance.

2.4 3D Isometric Workflow Replay

For learners and technical teams seeking a deeper conceptual understanding of model tuning, ProtoFlow provides a dedicated 3D isometric training replay view. It deconstructs the 7 fundamental stages of LoRA fine-tuning into an interactive 3D visual walkthrough:

Stage-by-Stage Deconstruction: Visually demonstrates data preparation, forward propagation, loss calculation, backpropagation, and low-rank matrix updates.
Interactive Playback Controls: Step through, replay from beginning, pause, and inspect the mechanics of parameter-efficient fine-tuning without complex math barriers.

3. Target Users and Use Cases

3.1 Target User Personas

Enterprise Decision-Makers & Business Owners: Convert historical support logs, product whitepapers, or standard operating procedures into testable prototypes to evaluate project viability and resource planning.
AI & Engineering Teams: Accelerate validation of dataset quality, model candidates, and hyperparameter sets while eliminating repetitive environment bootstrapping.
Training & Educational Institutions: Demonstrate LLM fine-tuning and parameter-efficient mechanisms through intuitive 3D simulations without requiring complex CUDA installations.
Individual Developers & AI Enthusiasts: Gain an out-of-the-box sandbox to explore fine-tuning techniques and align small models on personal hardware.

3.2 Typical Business Scenarios

Scenario	What Can Be Validated
Enterprise Knowledge Assistant	Train models on product manuals, internal regulations, and policy guides to evaluate factual adherence to business knowledge.
Customer Support Alignment	Fine-tune on historical inquiries and standard scripts to compare tone, phrasing adherence, and resolution rates.
Specialized Document Q&A	Synthesize Q&A pairs from complex technical documents to verify domain-specific comprehension and summarization.
AI Education & Hands-On Training	Explain LoRA principles visually and let students execute real end-to-end fine-tuning runs independently.

4. Technical Advantages and Deployment Options

4.1 Security and Identity Access

Multi-Factor Authentication: Supports traditional username/password authentication as well as enterprise OIDC-based Single Sign-On (SSO).
Hardened Handshake Protocols: Employs PKCE, state tokens, and nonces to protect against token interception and replay attacks.
Role-Based Access Control: Built-in Administrator and Regular User roles with strict API-level authorization checks.

4.2 Local and Elastic Cloud Inference

High-Performance Inference Engine: Supports containerized vLLM deployments for local low-latency inference.
Elastic Cloud Scaling: Seamlessly connects to cloud GPU instances to accommodate bursty training or inference demands.
Real-Time Streaming: Delivers typing-style responses via Server-Sent Events (SSE) for responsive user feedback.

4.3 Deployment Options

Deployment Mode	Ideal For	Key Characteristics
Local Single-Server	Individuals or small teams conducting rapid pilots	Consolidates frontend, backend, database, and training engine on a single server with mounted volumes.
Multi-User Server Deployment	Enterprise and institutional multi-user teams	Centralized backend and compute cluster with unified access control and resource quotas.
Cloud GPU Inference	Environments without local GPUs or needing elasticity	Local control plane dispatching inference workloads to on-demand cloud GPU instances.

[!NOTE] In private on-premises deployments, all business data, training jobs, and inference services remain securely within the enterprise network perimeter. Formal delivery specifications and hardware recommendations can be finalized during consultation.

5. Environment & Compatibility Recommendations

Recommended Display: 1920×1080 resolution or higher for optimal side-by-side comparison and 3D visualization rendering.
Supported Browsers: Latest stable releases of Google Chrome, Microsoft Edge, and Apple Safari.
Hardware Recommendations: Lightweight models (1.5B–7B) can be validated on a single 8GB–16GB VRAM GPU with appropriate quantization and LoRA configurations. Larger models benefit from multi-GPU workstations or cloud compute instances.

6. Technical Support and Services

To learn more about platform features, explore enterprise delivery options, or review implementation details, check our Frequently Asked Questions (FAQ) or reach out via official support channels.

LLM Security Gateway

Enterprise proxy protecting internal applications and customer chats from prompt injection, PII leakage, and compliance penalties.

1. Product Positioning

LLM Security Gateway is a unified security entry point and governance console for enterprise LLM access. It sits between business applications, developer platforms, and model providers, and centralizes model access, organization and project ownership, API key management, quotas and costs, security policies, intent detection, request auditing, and operational analytics.

The product does not replace the underlying models. Its purpose is to provide a manageable, auditable, and traceable control plane when multiple teams, model providers, and applications share LLM capabilities, helping enterprises transition model invocation from “scattered access” to “centralized governance.”

Core value	Description
Unified access	Bring providers, models, and calling entry points into one configuration and observability layer
Access governance	Define calling boundaries through organizations, projects, users, roles, API keys, and client sources
Security review	Correlate security checks, intent detection, policy hits, and request context in one audit path
Cost visibility	Use token, request, cost, budget, and quota views to support resource management
Troubleshooting	Investigate abnormal requests through request_id, model, channel, user, time, and status evidence

2. Business Context and Problems Solved

Enterprises adopting multiple LLM services often face fragmented key management, inconsistent calling conventions, unclear cost attribution, security findings without context, difficult failure investigation, and duplicated governance work across teams. LLM Security Gateway addresses these issues through a unified console and a consistent audit data path.

Business issue	Product response
Model-provider configuration is fragmented	Centralized management of channels, model catalog, provider status, and pricing information
API keys are hard to map to teams or projects	Organizations, projects, users, roles, and API keys establish clear ownership
Security findings lack reviewable evidence	Security, intent, and request audits show policy, risk, user, source, and request context
Token usage and cost are hard to control	Quota, budget, cost trend, and multidimensional usage views support management
Failures and unusual growth are hard to locate	Request logs, system logs, traces, and dashboards support investigation

3. Intended Users and Responsibilities

User role	Primary concerns	Common capabilities
Platform administrator	Model access, organization structure, permission boundaries, and system availability	Organizations, projects, models, channels, users, roles, API keys
Security and audit staff	Risky requests, policy hits, intent detection, and accountability	Security audit, intent audit, request audit, security traces, system logs
Operations staff	Usage trends, success rate, latency, abnormal growth, and activity	Operations overview, efficiency dashboard, model traffic, user activity
Finance or resource owner	Token usage, costs, budgets, quotas, and cost attribution	Quotas, billing, budget pools, cost trends
Application team or project owner	API key scope, call status, and project-level cost	Projects, API keys, request logs, model usage details

4. Operating Model

4.1 Request Governance Data Path

The product builds a consistent data path around each model request:

Business application / client
  -> API key
  -> Project
  -> Organization
  -> User / role
  -> Unified access layer
  -> Security policy / intent detection
  -> Model channel / model
  -> Request log / audit log / tokens and costs / operations dashboard

This path keeps organization, project, user, API key, request source, and detection result consistent across lists, details, filters, and export scopes where available. Administrators and audit personnel can start from any dimension and cross-check evidence through request_id, API key, project, user, or model information.

4.2 Core Objects

Object	Description	Main use
Organization	Enterprise or business-unit resource boundary	Contains projects, users, budgets, and governance scope
Project	Business system, application, or team workspace	Owns API keys, requests, costs, and quotas
User and role	Console operator and permission set	Controls page access, operations, and accountability
API key	Credential used by applications to call model services	Binds project, quota, status, and usage records
Client source	Calling application, client, or environment	Helps identify source, abnormal access, and ownership
Model channel	Connection to a provider or model service	Maintains base URL, status, probe result, and supported models
Model catalog	Callable models with capabilities, pricing, and tags	Supports selection, cost estimation, and traffic analysis
Security policy	Content-safety, intent, and access-control rules	Defines detection logic, risk levels, and handling basis
Request record	Status, time, model, and context for a single call	Supports tracing, troubleshooting, auditing, and analytics

5. Product Components

Component	Description
Administration console	Pages for organizations, projects, users, models, channels, API keys, security, audits, costs, and operations
Unified access layer	Consistent model-calling entry point for applications; exact APIs depend on the deployed edition
Governance capabilities	Identity, permissions, quotas, budgets, cost views, model channels, and project ownership
Security capabilities	Security policies, rules, templates, security models, security agents, intent detection, and risk review
Observability	Request logs, audit logs, traces, tokens, costs, success rate, and trend analytics

6. Core Capabilities

6.1 Organization, Project, and Access Governance

Use organizations and projects to hold users, API keys, budgets, quotas, and request ownership;
Manage users, roles, and permission scope for console access and operations;
Create, inspect, archive, and manage API keys, including their status and usage;
Keep organization, project, user, and key relationships consistent across request, audit, and cost views.

6.2 Model Channels and Model Catalog

Maintain model providers, base URLs, supported models, channel status, and probe information;
Manage model metadata, capability tags, context capabilities, pricing, and daily traffic trends;
Review channel relationships, abnormal states, and usage distribution in model details;
Reduce configuration and troubleshooting effort when multiple providers and models are used in parallel.

6.3 Security Protection and Intent Detection

Configure security policies, rules, templates, security models, and security agents;
Review security events, risk levels, policy hits, intent categories, and handling status;
Correlate detection results with request context, user, model, channel, and source;
Provide evidence for security review, false-positive analysis, and policy improvement.

6.4 Request Audit, Tracing, and System Logs

Search by time, status, model, channel, organization, project, user, API key, and risk dimension;
Use request_id to correlate request details, detection results, processing paths, and system logs;
Review the same event through request audit, security audit, intent audit, and security traces;
Locate failed requests, abnormal latency, concentrated errors, policy blocks, and source anomalies.

6.5 Quotas, Billing, and Operational Analytics

Review request, token, cost, latency, success-rate, and activity trends;
Analyze usage by organization, project, user, model, channel, and API key;
Manage resource control through quotas, budget pools, and billing views;
Identify high cost, low adoption, unusual growth, concentrated failures, and model traffic shifts.

7. Capability Matrix

Module	Key pages	Key indicators or fields	Use case
Operations overview	Operations Overview	Requests, tokens, costs, success rate, risk summary, trends	Give managers and operators a quick system-level view
Organization governance	Organizations, Projects	Organization status, project ownership, members, budget, resource scope	Define enterprise, team, and application governance boundaries
Model governance	Models, Channels	Provider, model, capability tags, pricing, channel status, probe result	Maintain multi-model access and investigate channel issues
Access governance	API Keys, Users, Roles	Key status, project ownership, permissions, quota, usage records	Control application entry points and permission scope
Quotas and costs	Quota, Billing	Tokens, costs, budgets, quota usage, cost distribution	Attribute cost, manage budget, and review abnormal spending
Security audit	Security Audit, Intent Audit	Risk level, policy hit, intent category, handling status	Review high-risk requests and improve policies
Request audit	Request Audit, Traces, System Logs	request_id, status, model, channel, user, source, duration	Troubleshooting, accountability, tracing, and evidence retention

8. Interface Examples

The following screenshots illustrate the information hierarchy and console interface style of the main pages.

8.1 Operations Overview

Operations overview example

The operations overview aggregates requests, tokens, costs, success rate, risks, and trends. It helps assess overall system health, recent changes, and areas that need deeper investigation.

8.2 Organization and Project Governance

Organizations example

Projects example

Organization and project pages maintain resource ownership. Projects serve as the key connection point between API keys, requests, quotas, costs, and audit evidence, establishing clear hierarchical relationships between organizations, projects, users, and keys.

8.3 Models and API Keys

Models example

API keys example

The model page shows the model catalog, provider relationships, capability tags, and traffic trends. The API key page shows credential status, project ownership, access scope, and usage information, making it a primary entry point for access governance and troubleshooting.

8.4 Quotas and Billing

Quota example

Billing example

Quota and billing pages are used to view budget pools, token usage, cost trends, and cost attribution.

8.5 Security, Intent, and Request Audits

Security audit example

Intent audit example

Request audit example

Audit pages are designed for security, operations, and platform administrators. Users can review requests by risk level, intent type, request status, model, channel, organization, project, user, and source, and open request details for supporting evidence.

9. Typical Use Cases

Use case	Operating path	Outcome
Connect a new business application	Create an organization or project, configure a model channel, create an API key, and set quota and permissions	The application receives a unified calling entry point while the platform retains ownership and audit data
Maintain model channels	Review the model catalog, channel status, probe results, pricing, and traffic trends	Administrators understand provider availability and model usage
Review a high-risk request	Start from security audit or intent audit, filter by risk level, then inspect request details and context	Security staff obtain evidence about policy hits, user, source, and handling status
Investigate abnormal cost	Locate cost movement from billing or the efficiency dashboard, then drill down by project, model, user, or key	Resource owners confirm the source of spending and adjust quota or budget
Troubleshoot call failures	Search request audit by status, time, model, or request_id, then correlate system logs	Operators identify failure scope, channel status, and affected applications

10. Environment and Compatibility

Recommended resolution: 1920x1080;
Target browsers: latest stable versions of Chrome, Edge, Safari;
Production deployment supports containerization, on-premises private network deployment, and other enterprise environment options based on the delivery plan.

11. Data Security and Compliance

Security policies, intent detection, and content inspection capabilities must be configured and validated against actual models, rules, and business scenarios;
Production data retention, access control, log masking, and audit export scopes are subject to the delivery plan and compliance requirements;
Cost, budget, and token metrics are used for usage analysis and resource governance, and actual billing is subject to enterprise procurement contracts and service provider standards.

12. Support and Services

For product details, technical support, or enterprise delivery inquiries, please contact our official support team.

Frequently Asked Questions (FAQ)

Frequently asked questions regarding products, deployment scopes, and technical support.

LLM Visual Training (ProtoFlow)

1. Do I need an algorithm background?

No prior experience with training scripts is required. Visual workflows, templates, and training monitoring lower the barrier for an initial validation; model selection and result evaluation should still be completed jointly by business and technical staff.

2. Can data remain within the internal network?

Yes. In a private deployment, data, training, and inference can run within the enterprise environment. The specific network and access-control design should be confirmed for the deployment environment.

3. What GPU is required?

It depends on the model, quantization method, sequence length, and batch settings. Lightweight models can be tried in an 8 GB VRAM environment when the appropriate quantization and LoRA settings are used; cloud GPUs can also be used according to the deployment configuration.

4. Can I build a dataset from documents?

Yes. The platform supports importing common document formats, parsing them, and generating Q&A datasets. Specific supported formats depend on the deployed version.

5. Which models are supported?

The platform includes models from the Qwen3 and DeepSeek-R1 families. The available model list depends on the deployed version and available compute resources.

Contact Us

If you have any questions about the documentation or need assistance, please contact us:

Official Email: contact@maxso.ai
Online Demo: Visit the ProtoFlow Static Demo for an instant preview of the product interface and workflow.