Project Intern: Autonomous Local Chief of Staff

Intern is a locally hosted, highly autonomous AI agent designed to act as a hyper-personalized Chief of Staff. It operates entirely on local hardware, processing sensitive daily data streams (calendar, financial, medical, and search history) to generate actionable micro-movements and life-management insights.

Because this agent handles highly sensitive personal data, the architecture is built on a foundation of absolute zero-trust security, bare-metal isolation, and hard API throttles.

🛡️ Core Security Architecture

This is not a standard API wrapper. The Intern is physically and logically boxed in:

1. Hypervisor Isolation (Proxmox KVM): The agent runs inside an isolated Ubuntu KVM virtual machine. No shared kernels.
2. Network Micro-Segmentation (SDN): The VM sits on a host-only Virtual Network (VNet). The default egress policy is DROP. The agent can only reach explicitly whitelisted endpoints (e.g., SMTP/IMAP).
3. Zero-Trust Access: No open router ports. Remote access is handled exclusively via a Tailscale SSH tunnel installed inside the VM.
4. Ephemeral Code Execution: The agent dynamically writes Python to solve tasks. This code is NEVER executed on the host OS. It is trapped inside a disposable, unprivileged Alpine Docker container (Dockerfile.burner) that is destroyed the moment the code finishes running.
5. The Kill Switch: A LiteLLM Proxy sits between the agent and the local inference engine, enforcing strict Tokens-Per-Minute (TPM) and Requests-Per-Minute (RPM) limits to prevent infinite logic loops and hardware meltdowns.

🧠 Memory & State Management

The Intern uses a multi-tiered memory architecture designed to survive model switches and context-window compaction:

• The Boot Sequence (AGENTS.md): Loaded immediately upon instantiation. Forces the agent to read its constraints, ingest daily data, and review past mistakes before responding to the user.
• The Handover State (memory/YYYY-MM-DD.md): Append-only daily logs used to preserve context across long sessions.
• The Operations Manual (LEARNINGS.md): A strict record of the agent's past failures and the rules it generated to never repeat them.
• Long-Term Storage (agent_memory.h5): Powered by EdgeHDF5, this provides lightning-fast, zero-daemon Hybrid Search (Vector + BM25) for recalling exact variables, names, and concepts from past sessions.

📥 The Secure Data Ingest Pipeline

To analyze life data without exposing it to cloud APIs, we use a secure "Drop Box" methodology.

The agent has NO ability to scrape the user's personal devices. Instead, automated, encrypted cron jobs on the user's primary devices push daily text streams (JSON/CSV/TXT) through the Tailscale tunnel directly into the workspace/ingest/ directory.

During the Boot Sequence, the Intern parses:

• calendar_events.json
• search_history_72h.txt
• financial_alerts.csv
• health_wellness_log.txt

It synthesizes this data into an actionable Boot Report, identifying scheduling conflicts, financial anomalies, and actionable steps toward the user's overarching goals, then deletes the ingest files to maintain a clean state.

🛠️ Tech Stack

• Framework: smolagents (CodeAgent)
• Local Inference: Ollama (Qwen 2.5 Coder 7B)
• API Gateway: LiteLLM Proxy
• Execution Sandbox: Docker (Python Alpine)
• Vector Backend: EdgeHDF5

🚀 Getting Started

(Setup instructions for provisioning the Proxmox KVM, configuring the Proxmox SDN firewall, and launching the systemd services are detailed in /docs/project_intern_adr.md)