ZDXAI

Local-First AI That Knows You

Private. Intelligent. Always On.

Executive Summary

ZDXAI is a local-first, on-device AI assistant built for people who refuse to compromise on privacy. No cloud. No API keys. No data leaving your machine. Every inference runs on your hardware using quantized GGUF models, delivering Jarvis-level capability without surveillance-level exposure.

Built by ZeroDriveX LLC, ZDXAI combines a multi-model architecture with Retrieval-Augmented Generation (RAG), MCP tool execution, LoRA fine-tuning, and an intelligent Embedding Gravity Well routing system that learns your usage patterns over time.

This is not another ChatGPT wrapper. This is your personal AI that runs entirely under your control, on your infrastructure, with your data staying exactly where it belongs: with you.

The Problem With Cloud AI

Every major AI assistant in production today has the same fundamental architecture: your data flows to someone else's servers, gets processed by models you don't control, and lives in logs you can't audit. For casual use this is an acceptable trade-off. For professionals, developers, and security-conscious users, it is not.

What You Lose With Cloud AI

• Privacy: Conversations, documents, and context are processed on third-party infrastructure
• Control: Model behavior, updates, and capabilities are determined by the vendor
• Cost predictability: Token-based pricing creates unpredictable monthly bills
• Offline capability: No internet means no AI assistant
• Customization: You cannot fine-tune a hosted model on your own data
• Security: Your intellectual property passes through systems you do not own

ZDXAI was built to eliminate every one of these trade-offs without sacrificing capability.

System Architecture

ZDXAI is built around four interconnected systems that work together to deliver intelligent, context-aware responses while keeping everything local.

Multi-Model Architecture

ZDXAI auto-selects the optimal model based on available hardware resources:

Models are served as quantized GGUF files with GPU offload support via --gpu-layers for dramatically faster inference on systems with compatible hardware.

Embedding Gravity Wells

The most architecturally novel component of ZDXAI is the Embedding Gravity Well routing system. Rather than sending every query to the same model, ZDXAI learns from your usage patterns over time to steer queries toward whichever model or platform will handle them best.

This is a learned routing system. The more you use ZDXAI, the better it understands which type of query benefits from which resource — local fast model, larger local model, or an optional cloud API call when local capability is insufficient.

Retrieval-Augmented Generation (RAG)

ZDXAI can ingest your documents and use them as grounded context when answering questions. This transforms ZDXAI from a general assistant into a domain expert on your specific knowledge base.

Supported file types: .txt, .md, .json, .jsonl, .csv, .pdf

The RAG pipeline chunks documents, creates embeddings, and uses cosine similarity search to surface the most relevant context before generating any response. Memory is searchable and persistent across sessions.

MCP Tool Execution

ZDXAI can take action on your system through a built-in Model Context Protocol tool system. All tools use a deny-by-default permission model — ZDXAI will ask for your approval on first use of any tool and respect whatever permission level you set.

Permissions are persistent and user-controlled. Set filesystem.read to allow, shell.execute to ask, and web.search to deny — ZDXAI respects those settings until you change them.

LoRA Fine-Tuning

ZDXAI includes built-in LoRA adapter training, allowing you to specialize the base model for your specific use case without requiring expensive GPU infrastructure or cloud training pipelines.

Three Training Modes

• Conversation training: Use your own chat history as training data with /train
• Document training: Train on ingested documents with /train-doc
• JSONL training: Provide a custom labeled dataset with /train-file

Training uses standard JSONL format with instruction, input, and output fields — the same format used by the broader open-source fine-tuning ecosystem. This means datasets prepared for other tools work directly with ZDXAI.

Adapters are hot-swappable. Activate a domain-specific adapter for specialized work, switch it off to return to the base model — no restart required.

Workflow Automation

ZDXAI supports multi-step workflow automation by chaining tools together in sequence. Workflows allow you to define repeatable automated processes that ZDXAI executes end-to-end.

Built-in workflow examples include system health checks (run diagnostics across three tools in sequence), search-and-fetch pipelines (web search then fetch the top result), and project overviews (list directory, read README, summarize).

Workflows accept runtime parameters and run all steps in order, reporting success or failure at each stage. This is the foundation for building Jarvis-style automation sequences tailored to your workflow.

Privacy & Security Model

ZDXAI was designed with a security-first architecture from the ground up. Every decision in the system reflects a bias toward keeping your data local and your control absolute.

Core Privacy Guarantees

• All inference runs locally on quantized GGUF models — nothing leaves your device by default
• No telemetry, no usage tracking, no call-home behavior
• No API keys required for standard operation
• No accounts, no subscriptions, no vendor relationship with your data
• Conversation history stored locally and fully under your control

Deny-By-Default Tool Permissions

The MCP tool system enforces deny-by-default on every tool. When ZDXAI determines that a tool would help answer your query, it presents the specific operation for your approval before executing. You can approve once, always, or deny permanently per tool.

This architecture ensures that ZDXAI cannot take action on your system without your explicit knowledge and consent — a meaningful security property that cloud-based AI assistants cannot offer by design.

Multi-Platform Cloud Fallback (Optional)

When local models are insufficient for a complex task, ZDXAI optionally integrates with OpenAI, Anthropic, Gemini, and xAI APIs. This is strictly opt-in and requires explicit API key configuration. No data is sent to cloud APIs without user-initiated configuration.

ZDX Guard Integration

ZDXAI is designed to operate alongside ZDX Guard, ZeroDriveX's prompt injection detection and sanitization layer. Where ZDXAI provides the intelligence layer, ZDX Guard provides the security layer.

Every piece of text entering ZDXAI from external sources — documents, web fetches, pasted credentials, API responses — can be routed through ZDX Guard's detection engine before being processed by any model. This closes the supply chain attack vector that affects every other AI assistant in production.

Use Cases

Developer Productivity

• Ingest your entire codebase as RAG context and query it conversationally
• Debug code with an AI that has full filesystem access under your control
• Fine-tune on your coding style and patterns for personalized suggestions
• Run automated workflows for project health checks and documentation generation

Security Professionals

• Process sensitive reports and findings without cloud exposure
• Query threat intelligence documents with RAG without sending data to third parties
• Use ZDX Guard integration for prompt injection protection on all model interactions
• Run fully air-gapped on isolated infrastructure

Knowledge Workers

• Ingest research papers, meeting notes, and project specs as searchable memory
• Resume conversations with full context via conversation history and ID-based resumption
• Train domain-specific adapters from your own document library
• Automate recurring research and summarization workflows

Roadmap

Near-Term

• Web UI and mobile interface for non-terminal users
• ZDXAI + ZDX Guard unified security stack deployment
• Expanded model support including Phi, Qwen2.5, and larger quantized variants
• GUI adapter management and training progress visualization

Medium-Term

• Voice interface and audio transcription pipeline
• Agent orchestration for multi-step autonomous task execution
• Team deployment mode with shared RAG knowledge bases
• Enterprise MDM integration for fleet deployment

Long-Term

• The fully realized ZeroDriveX AI stack: ZDXAI + ZDX Guard + fine-tuned injection detection model
• Private federated learning across user instances with zero data sharing
• Full Jarvis-style ambient assistant with always-on awareness and proactive suggestions

Getting Started

ZDXAI is available now on GitHub. Installation takes under five minutes on any Linux, macOS, Windows, or Android (Termux) system.

Full documentation, GPU acceleration setup, and troubleshooting guidance is available in INSTALL.md in the repository.

About ZeroDriveX LLC

ZeroDriveX LLC is a Wyoming-based cybersecurity and AI infrastructure company founded in 2024. The company builds privacy-first developer tools, security products, and AI systems designed for users who require control over their own data and infrastructure.

Products include ZDXAI (local AI assistant), ZDX Guard (LLM prompt injection protection), AUTH1 (JWT/Redis authentication infrastructure), and ZDX Mobile AI (Ollama compiled for Android).

zerodrivex.com