DAX

AI & Agent Engineer

Building local-first AI agents, developer tools, and safe computer automation.

“I build agents that do more than generate text — they plan, act, verify, and improve.”

  🧠 Intelligence               🛠️ Tools                    🛡️ Verification
  Local Multi-Backend           Sandboxed Execution         Evidence-Backed Results
  Task Planning & Memory        AST Code Manipulation       Human Approval Gating

We don't just build chatbots. We build systems where artificial intelligence understands an objective, coordinates real-world tools, executes work, and validates the result before declaring completion.

Our engineering focus is dedicated to:

• Autonomous AI Agents: Structured reasoning loops that combine memory, decomposition, and execution.
• Local AI & Privacy: Keeping data, prompts, and inference strictly on local hardware.
• Coding Automation & AST Verification: Moving beyond raw text generation to syntax-verified diff patching.
• Computer Interaction: Safe, sandboxed desktop action engines bound by strict operating system guardrails.
• Deterministic Tooling: Wrapping stochastic models with reliable schemas, bounds, and audit trails.

Artificial intelligence becomes truly valuable when it transitions from:

“Telling you what to do” ──► “Understanding the goal, executing the work, and verifying the result.”

Modern agent engineering shouldn't depend solely on brittle prompt instructions or opaque cloud runtimes. By focusing on local execution, deterministic tools, human-in-the-loop control, and empirical verification, we can build autonomous software that is powerful, transparent, and completely under user control.

Our work is guided by five core principles:

1. Local First — Whenever possible, reasoning, context, and data stay on the local machine.
2. Tools, Not Just Prompts — Models become practically useful when equipped with structured, reliable tools.
3. Verify, Don't Assume — Successful execution must always be backed by evidence (AST parsing, tests, dry-runs).
4. Human Control — Sensitive, destructive, or state-altering actions require explicit human confirmation.
5. Build Small, Compose Systems — Small, decoupled, and thoroughly tested components compose into robust agent architectures.

                                  +-----------------------------+
                                  |     USER GOAL & CONTEXT     |
                                  +--------------+--------------+
                                                 |
                                                 v
                      +----------------------------------------------------+
                      |     FLAGSHIP AGENT RUNTIME: nexus-agent            |
                      |  • Dynamic Task Planner    • Local Model Router    |
                      |  • Two-Tier Memory Store   • Safety Policy Engine  |
                      +--------------------------+-------------------------+
                                                 |
                                [ Dispatches To Specialized Tools ]
                                                 |
                     +---------------------------+---------------------------+
                     |                                                       |
                     v                                                       v
+------------------------------------------+    +------------------------------------------+
|      CODE TOOLING: micro-coding-agent    |    |   DESKTOP TOOLING: desktop-action-agent  |
| • Safe Workspace Reader                  |    | • Window Discovery & Focus Manager       |
| • Pre-Flight Unified Diff Dry-Runs       |    | • Process Whitelist & Command Filter     |
| • Human-in-the-Loop Approval Gating      |    | • Bounded Input Simulation (Click/Type)  |
| • AST Syntax Integrity Verification      |    | • Cryptographic SHA-256 Audit Trail      |
+------------------------------------------+    +------------------------------------------+

Local-First AI Agent Reference Core
Dynamic model routing (Ollama / vLLM / mock), task decomposition, 2-tier context & episodic memory, and bounded workspace tools.
Status: Public • Tests: 7/7 Passing • CI: Passing • Explore Repository →

Deterministic AI Coding Engine
Workspace indexing, atomic diff patching, pre-flight AST syntax verification, and approval gating.
Status: Public • Tests: 12/12 Passing • CI: Passing • Explore Repository →

Safe Sandboxed Desktop Automation
Window management, application allowlists, bounded screen inputs, and tamper-evident SHA-256 audit logs.
Status: Public • Tests: 10/10 Passing • CI: Passing • Explore Repository →

• Local Agent Runtimes: Low-latency orchestration against local quantization runners.
• Computer-Use Systems: Bounded OS interaction primitives with accessibility tree integration.
• Verification-Driven Coding: Automated test generation, AST diagnostics, and rollback engines.
• Agent Memory Topologies: Hybrid episodic buffers and vector retrieval for long-horizon tasks.
• Safe Autonomous Execution: Cryptographic audit trails and fine-grained capability isolation.

• Core & Runtime: Python 3.9+, Bash, Subprocess Management
• AI & Reasoning: Local Inference (Ollama / vLLM / LM Studio), Tool Calling, JSON Schema Validation
• Code & Parsing: Abstract Syntax Trees (ast), Unified Diff Engines (difflib)
• Safety & Verification: Sandboxing, Permission Policies, SHA-256 Cryptographic Audit Logging
• Quality & CI: Pytest, GitHub Actions Multi-Version Matrix, Ruff, Mypy

These repositories are independent reference implementations designed to explore and document practical agent engineering patterns. Every project includes real automated test suites, clean documentation, and MIT licensing.

• LinkedIn: Mohamed Musa Ahmed
• GitHub: @dax0056
• Security & Inquiries: dax0056@users.noreply.github.com