Kami Technology OS
AI-Powered Personal Operations
An intelligent operating system that connects your email, messaging, finances, and knowledge into one autonomous platform. Built for a single operator. Engineered to handle everything else.
Capabilities
Six subsystems, one platform
Each domain is a complete system on its own. Together they form a single coherent operating layer across your digital life.
Intelligent Email
AI classification, declarative filters, automated sweeps, and push notifications across multiple accounts.
Autonomous Agents
OODA-loop agents that research, plan, and execute multi-step tasks, reporting findings incrementally.
Financial Oversight
Consolidated multi-account tracking with daily snapshots, monthly reports, and automated invoice extraction.
Communication Hub
Telegram, Discord, and Gmail unified with continuous backup, full-text search, and AI-generated summaries.
Knowledge & Memory
Personal library, structured categories, encrypted vault, and persistent memory across conversations.
Always Watching
Health monitoring, incident detection, automatic escalation, and connector sync across dozens of services.
Intelligent Email
Every incoming email is classified by an LLM that reads the headers and snippet, assigning labels — urgent, requires reaction, personal, close circle, clutter — so you see what matters without reading everything. The close-circle system tracks contacts across email addresses, Telegram handles, and Discord identities, ensuring messages from the people who matter are never buried.
Filters are managed declaratively: define what should exist in configuration and the system synchronises to Gmail, creating, updating, or removing server-side rules as needed. Periodic cleanup runs on filters that accumulate noise — an after-age rule deletes old matches, while an after-event rule uses the LLM to extract event dates from messages and cleans up once the event has passed.
On-demand insights answer questions like 'what needs my attention?', 'anything suspicious this week?', or 'show me stats'. These are LLM-generated analyses, not canned reports — tailored to the question being asked, drawn from real correspondence.
Autonomous Agents
Errand boys are autonomous agents spawned to work on specific tasks. Each receives the errand's context — descriptions, documents, links, notes — and a brief, then works independently through the OODA cycle: observe the situation, orient against the brief, decide the next action, act, and repeat. Findings arrive incrementally via a bulletin board, not as a single dump at the end.
Every agent operates within strict guardrails: a maximum number of iterations, a time limit, a web-search budget, and an LLM cost ceiling. When any limit is reached, the agent stops and reports what it has. Certain actions — form submissions, payment decisions — require explicit human approval before proceeding.
For larger questions, the research system deploys a HADI model: a Lead agent formulates testable hypotheses, Field agents execute them in parallel, and the Lead aggregates findings across cycles — converging towards an answer rather than expanding into ever-wider search.
Financial Oversight
Every morning, connectors pull balances from bank accounts, credit cards, crypto wallets, and investment platforms. The snapshot engine converts everything to a home currency using ECB and CoinGecko rates, then stores the result. If a connector fails, the snapshot uses the last known balance and marks it stale — configurable thresholds trigger alerts if staleness exceeds acceptable limits.
Monthly reports capture the financial position over a period: grouped sections with subtotals, net position summaries, rendered in HTML, PDF, CSV, or XLSX. Reports can be published to an external git repository for archival outside the system. Invoices are extracted from email automatically — matched to known services by sender pattern or LLM fuzzy matching.
Running costs track what the system itself spends — LLM API calls, hosting, database usage, search queries — broken down by provider and by day, so you always know what your infrastructure is costing.
Communication Hub
Telegram messages are backed up continuously through the client API — the same protocol the app uses — giving access to all chats: private conversations, groups, supergroups, channels, and forum topics. Discord messages are captured through the gateway WebSocket. Both support multiple accounts with per-chat backup policies controlling what gets captured, summarised, or ignored.
The transport layer unifies outbound messaging across platforms with four urgency levels: authorisation (2FA codes, immediate delivery), urgent (security alerts, critical failures), push (unexpected situations), and chat (conversational, no push notifications). A single channel can map to multiple platforms simultaneously for redundancy.
Chat insights analyse backed-up messages: activity statistics, unanswered conversations, per-chat summaries with mention detection, and contacts who have gone silent — with estimated typical frequencies so you notice when someone drops off.
Knowledge & Memory
The library manages a personal collection of books — EPUB, PDF, FB2, MOBI, DJVU — with metadata extraction, cover thumbnails, and full-text indexing. New books can be acquired through a Telegram bot integration that locates, downloads, and imports them into the collection automatically. External repositories, Dropbox, and Notion workspaces are connected for file access, document editing, and knowledge organisation.
The vault stores structured personal data: identity documents, tax registrations, bank details, contact information. Each entry carries a category and country, deliberately simple — a place to keep the information you need twice a year and can never find when you do. Categories provide a unified classification layer across all features, tracking entities (contacts, services, locations) with consistent identity resolution.
Conversation memory operates at two levels: channel memory for specific threads, global memory across all conversations. Both are subject to compaction — older entries are summarised while pinned facts stay exactly as written until explicitly removed.
Always Watching
Connectors run on a schedule — API connectors for services with proper interfaces, Playwright connectors that automate browsers for everything else, email connectors that parse data from messages, and Telegram bot connectors that interact with service bots. Every connector produces normalised data conforming to a predefined schema, so a bank balance from any provider ends up in the same structure.
When something goes wrong — a connector fails repeatedly, credentials expire, a service becomes unreachable — the system creates an escalation: a persistent alert that lives in the database until explicitly dealt with. Unlike notifications that can be dismissed from a lock screen, escalations have a lifecycle (active, dismissed, resolved) and demand resolution. Deduplication prevents alert storms: same-source alerts within a window are merged, and severity can only escalate upward.
The platform runs health checks across all subsystems. Incidents are detected, analysed, and classified automatically. The dashboard surfaces what needs attention; silence is the most common outcome, and the desired one.
Process
How it works
Connect
Link your services. Email, messaging, finance, storage, cloud infrastructure — all authenticated and encrypted.
Configure
Define automations, agent behaviours, classification rules, and notification preferences. Everything is declarative YAML.
Delegate
The platform handles the routine. Agents work autonomously, summaries surface what matters, and decisions stay with you.
Built for one. Engineered for everything.
Kami Technology OS — personal infrastructure, not another SaaS subscription.