> Markdown version of https://etgarbonar.com/ai/ > The HTML page is the canonical source. Title: A private AI system, running under real load | Etgar Bonar Description: Not a demo. A private AI system running under real load, with permission boundaries, human approval, watchdogs, backups and a written log of every failure. Evolving since February 2026 # A private AI system: one builder, many agents, real consequences. It remembers from sources, connects small-business workflows, and acts only within explicit permissions. It runs across platforms, swaps models, and lives on hardware that cost less than a laptop. [The point of view it shaped](https://etgarbonar.com/serviceable-line/) How this site is made ## AI agents made this site. The framework, the judgement calls and the voice are mine. Since 30 August, every page here has been produced with AI agents I run, mainly Claude Code and Codex, working inside rules and review steps I set: the research, the drafts, the code, the [diagrams](https://etgarbonar.com/serviceable-line/shapes/), a [four-minute explainer video](https://etgarbonar.com/serviceable-line/#watch) and the LinkedIn carousel. I choose the argument, decide which words survive, and sign off on what goes live. **Every number is sourced** Each figure on this site traces to a ledger of sourced claims. A number without a source does not ship. **Private material cannot ship** Every release passes an automated gate that refuses anything carrying private notes or internal references. **Copy is reviewed line by line** Each page is mirrored into a document I comment on, and my comments become the next draft. **Big calls are argued first** Decisions go to a panel of advisor personas that disagree with each other before I decide. This section came out of one. One thing that went wrong In early October, a file on this site was found listing the paths of my private notes, though not their contents. The release gate above exists because of it. Seen in the system ## It remembers from sources, and the sources are connected. Facts come from real notes, records and logs rather than from conversational memory. Each answer stays traceable to the file it came from, and the files reference one another. Every bar in these captures is a deliberate redaction, not a rendering fault. What gets withheld follows a written disclosure rule, decided once, rather than a judgement call made page by page. Live knowledge map Recorded from the working system *Video: A live knowledge graph zooming from the whole collection into one selected page and its connections.* Real system recording 1. **The whole knowledge base** Every dot is a note. Dense clusters are subjects with deep, connected material behind them. 2. **Zoom into the relevant neighbourhood** Moving closer makes individual pages legible while keeping the relationships around them visible. 3. **One page, with its connections intact** Selecting a page reveals the frameworks, plans and references around it, so an answer remains traceable. **View the original still captures**  One note and the sources connected to it.  The full knowledge map. Where it works ## One operating layer for work and life. Business work Commercial strategy, AI visibility, product positioning, and decision support for confidential advisory work. Creative and web Websites, diagrams, explainer videos and carousels, built by agents and published after my review. See how this site is made. Personal administration Inbox, calendar, tasks, documents, decisions, and project continuity. Family and travel Shared planning, groceries, trips, chores, allowance, and family coordination. Household and finance Vendors, receipts, payments, shared expenses, and recurring follow-ups. Connected home Lights, scenes, sensors, cleaning, watering, alerts, and household routines. Health Health records, test results, and history kept connected, organized, and searchable. Operational evidence ## Built for real work, and run with the discipline it requires. **Applied capability** - Supports confidential advisory work in commercial strategy, positioning, AI visibility, and decision support - Daily use through WhatsApp and Telegram, in English and Hebrew - Source-backed memory plus connected tools and home systems **Reliability** - Nightly deterministic self-audits and canary checks - Fallbacks across AI providers, with outage and recovery alerts - Nightly offsite backups, a recovery runbook, and an incident log - Watchers detect critical file replacement, session failures, and intrusion events, repair recoverable states, and alert when attention is needed **Governance** - Specialists receive only the tools their job needs - The chores agent cannot inspect family photos because image access is not available to it - Long-term memory, usage, and cost are reviewed on a defined rhythm Multiple models, one governed system ### Model choice stays flexible. Usage stays visible. Requests can move between providers when needed, while usage and cost remain measured by provider and every fallback stays visible. Subscriptions carry the work. Metered API billing is used only where no plan covers the job, which keeps it a deliberate exception instead of the default.  **Plan consumption is visible too.** Session and weekly limits are tracked before they become blockers. AI usage Daily report Estimated model cost today **Cost is a governed number, not a surprise.** Daily spend is measured against a trailing average, and a provider fallback remains visible when it fires. **View the original redacted report**  Seen in the system ## It runs the house it lives in. The connected home is part of the same operating layer, not a separate hobby. This real phone walkthrough moves from the whole house to one room, a shared occasion and the watering system underneath it. Live home control Recorded from the working system *Video: A real phone walkthrough moving from a house overview to room controls, entertaining mode and the watering dashboard.* Real phone recording 1. **The whole house at a glance** Rooms and devices are grouped by floor, so the system understands where an instruction belongs. 2. **One room, direct control** Lights, blinds, climate and media remain available in one place when direct control is useful. 3. **A moment, not a device list** Entertaining brings the living room, kitchen, entrance, staircase and gate into one useful view. 4. **The infrastructure underneath** Watering totals, valve state, flow and missed-run alerts are visible when the family needs the detail. **View the original redacted home captures**  Four levels, grouped into one readable map.  Shared-space controls grouped around entertaining. Watching and raising the flag when needed ### The interface is only half the point. The system confirms scheduled work, keeps reading the sensors and raises an exception without waiting for someone to inspect a dashboard. Garden monitor Automatic **An alert nobody asked for.** It confirms the work, notices an abnormal reading and sends the exception to the family. **View the original redacted alert**  Works where the conversation is already happening ## It coordinates travel and stays with the group. There is no new app to learn. The assistant joins the WhatsApp group already in use, so questions, updates and shared decisions stay visible to everyone. It answers questions, writes briefings and keeps watching the bookings. Choose a real trip story **Flight change: A booking changed** **Trip planning** WhatsApp group What happened behind the message **A question: How hard is the climb?** **Trip planning** WhatsApp group What happened behind the message **Tomorrow's brief: Sent before anyone asked** **Trip planning** WhatsApp group What happened behind the message Recreated from real exchanges. Names and personal details have been removed. **Open the original, redacted WhatsApp chats from the trip group**    Works where the conversation is already happening ## Shared ledgers for chores, budgets, and tasks. People report chores in the WhatsApp group they already share, where everyone can participate and see the result. The price list is fixed in advance, so what a job is worth is not renegotiated at the end of the week. **Shared chores** WhatsApp group What happened behind the conversation **Pocket money, and an assistant that refuses to guess.** The familiar chat is only the surface. The important part is the rule-following behind it: verify the wording, ask when uncertain, then update the ledger deterministically. Recreated from a real group exchange. Names and personal details have been removed. **Open the original, redacted WhatsApp chat from the shared chores group**     Seen in the system ## It watches itself, and says so when something is wrong. A health panel regenerates through the day across both the knowledge base and the machine underneath it. Alerts fire only on red, plus one message when a fault clears, so silence is meaningful.  **The health panel, mid-day.** Backup freshness, credential checks, scheduled jobs, disk, memory, load and connectivity stay visible in one report. Here, one service is red while the rest remains healthy. That is when a WhatsApp alert fires, followed by one recovery message when the fault clears. Trust by design ## Private enough to use. Inspectable enough to trust. The operating system and its readable knowledge base run on a private home server. External AI models are interchangeable services, not the owners of the system or its memory. Checks, incidents, and fixes remain visible. When a failure repeats, it is turned into a rule or a deterministic tool so the same problem is less likely to happen again. The idea ## Not another chatbot. A personal system that works where life already happens. This is a working pilot of a private operating system with long-term, source-backed memory. It connects knowledge and actions across the tools a person or family already uses, while keeping the system readable, inspectable, and portable. I built it for three reasons: to experience first-hand, as an early adopter, how buying changes when AI does the finding, choosing and using; to run my own and my family's work and life on it; and to build differently. I have been a builder many times before. What is new is that I can now write the code: connecting systems and apps that were never connected, and automating processes that could not be automated before. Together with solving discoverability and selectability at Lokalise, it changed how I think product-led and sales-led motions evolve in the age of AI. It is now the platform, a second brain and its agents, that I use to assess where a company's Serviceable Line sits. How it works ## Three ideas make it different. 01 ### It remembers from sources Important facts come from real notes, records, tasks, and logs. They remain traceable to their source instead of disappearing into vague conversational memory. 02 ### It coordinates across life Focused assistants connect personal administration, family and travel, household and finance, the connected home, and the flow of information between them. 03 ### It keeps the owner in control Access is scoped by task. Durable memory requires human approval, sensitive actions keep a confirmation step, and the knowledge remains readable outside any one AI model. Who is behind it ## Built and operated by [Etgar Bonar](https://etgarbonar.com/), a commercial executive who chose to learn AI by building and running a working system around real life. He also built a mobile development workspace so he can inspect, program, and improve the system from anywhere. ## Want to talk about the system? [Get in touch](mailto:ai@etgarbonar.com)