Focused AI services, not one exposed chatbot.
Kaizen is divided into runtime areas with different trust levels. Max is the private assistant. KaizenAI Open is the public assistant. Search, voice, real-time calls, memory, communications, reminders, automation, image generation, and orchestration remain separated so each path can be secured, logged, and tested. The current stack reports 33/33 services running across 33 registered service definitions.
Access Boundary
Assistant Layers
Each layer has a narrower job and a clearer security boundary.
Max
The private assistant layer for authenticated owner workflows, memory-aware conversations, iOS continuity, and model routing.
KaizenAI Open
The public assistant layer for limited general chat and search-grounded answers. It has no private memory and no internal system access.
Orchestrator
The management layer for service health, prompt editing, memory tools, public chat controls, audit output, and operational diagnostics.
Max Voice
The real-time call persona, tuned for spoken conversation: time-aware responses, speakable output instead of screen formatting, and local speech processing. Approved users get the full assistant; anyone else reaches a bounded public demo with no memory and a short daily limit. The operator holds the kill switch on both.
Current Runtime Setup
The active stack is organized around daily use, testing, observability, and controlled public access. It was recently split into separated service areas so each path can be secured, scaled, and tested on its own instead of one process doing everything.
Chat Runtime
Routes private and public conversations through the correct prompt, search, memory, and safety path. A front-door gateway now handles sign-in, identity, and access-tier checks before a separate execution area runs the conversation.
Search Runtime
Handles current web questions, source-aware summaries, weather and finance lookups, and response cleanup. It now runs as its own dedicated service so search can be tuned and tested without touching the chat path.
Voice Runtime
Supports speech input, speech output, runtime status messages, mobile voice workflows, and the live phone lane — a dedicated low-latency speech tier lets approved users hold a real-time spoken conversation with Max.
Public Runtime
Queues and rate-limits KaizenAI Open requests on its own isolated path while blocking private tools and personal memory. The recent split keeps the public lane structurally separate from private execution.
Current System Map
The public shape of Kaizen now matches the private operator stack without exposing private endpoints, prompts, or model internals.
Gateway / Runtime / Search
Separate front-door, execution, and current-context services keep sign-in, queueing, search, and generation policy easier to audit.
Memory / Qdrant
Private semantic recall is user scoped and managed through the protected dashboard, not sent to the public assistant lane.
Voice / IMGEN
Speech workflows and image generation run as controlled capability areas with status checks, routing rules, and operator visibility.
MCP / Cloud Lanes
External tools and governed cloud model lanes sit behind policy gates so private and public users receive only the capabilities assigned to them.
Deep Research
The deep-research lane runs as its own loopback service: multi-source sweeps verified in a bounded second pass, returning cited reports instead of confident guesses — private and auditable.
Health / Backups
A standing health collector, live probes the assistant can read, a backup center with scheduled sync and retention, and an encrypted local secret vault behind the operator boundary.
Communications & Reliability Agents
The July 2026 work added real messaging surfaces, profile-scoped access, and delivery proof instead of one-off notification scripts.
Communications Hub
Telnyx-backed inbound SMS, operational call alerts, outbound delivery receipts, contacts synced from Kaizen profiles, STOP/START/HELP handling, and dashboard traceability.
Max SMS
Short SMS replies use a hidden max:sms lane and the sender's canonical X-Kaizen-User identity so memory remains scoped to the right person. An unrecognized number is answered on the public demo lane instead — no memory, no stored profile, a short daily limit, and STOP honored immediately.
Reminder Agent
Natural-language reminders can be created, listed, and cancelled through Max, then delivered to known users through SMS, email, Pushover, and future call paths.
Call Agent
Real-time phone conversations with Max are live: an approved user calls the Kaizen number and talks naturally, with local speech processing, the same profile boundary as every private lane, and an operator kill switch.
Beacon Agent
A private scheduled-check pattern that records a source, transcribes locally, classifies the result, sends alerts, stores a SQLite ledger, and verifies delivery receipts.
Notification Router
Severity-aware alert routing to email and push channels with quiet windows, delivery history, and operator-tuned policy so real failures surface and noise does not.
Task Agents
Beyond conversation, Kaizen runs narrow working agents that each do one job well, with their own limits and audit trails.
Research Agent
Explicit deep-research jobs: local-first source discovery, bounded verification of claims, and a final report with stable citations — resumable, owner-scoped, never a silent background crawler.
Scouts
Metro-scoped event scouts sweep public sources for upcoming happenings and deliver a clean digest by email on schedule — a small agent that makes weekends better.
Media Import
A private, confirmation-gated capture lane: the operator approves every job before any network retrieval, progress is durable, and finished artifacts deliver straight to the iOS app.
TagGen
Local music-library housekeeping: fingerprint identification, metadata proposals, and cover-art retrieval with dry-run-first, approval-gated writes that never touch audio.
How a Request Flows
Every request — public or private — walks the same governed path, which is why one platform can safely carry many trust levels.
1. Identity first
The gateway authenticates the caller and resolves their access tier before anything executes. Anonymous traffic can only ever become KaizenAI Open traffic.
2. Persona and tools
The runtime picks the right persona and hands it only the tools that tier allows — search, memory, schedule reads, automation — nothing more, enforced server-side.
3. Governed execution
Heavy work queues through a concurrency governor, progress streams as durable jobs, and the result lands with logs, receipts, and history the operator can inspect later.
Access & Governance
Who can reach which capability is decided at the gateway, before a request ever reaches a model — not left to the client to enforce.
Access Tiers
Every user belongs to an access tier that decides which models, tools, and limits they get. The owner tier is unrestricted; trusted and guest tiers are progressively narrower, and the rules are applied centrally rather than trusted to each device.
Concurrency Control
A runtime governor caps how many heavy local requests run at once and queues the rest, keeping the platform responsive under load instead of overcommitting hardware.
Operator Dashboard
A protected control plane manages users and tiers, memory, conversation-sync health, public-chat limits, prompts, service status, and per-service logs in one place.
Development Agents
Kaizen also uses command-line AI agents for its own software work, audits, and documentation loops.
Code Operations
CLI agents can inspect source trees, patch services, run tests, compare diffs, and prepare controlled releases through the private development workflow. On the schedule platform, multiple agent vendors connect under a server-enforced proposer role: they can read and stage, and the platform refuses approve, apply, and export no matter which agent asks.
Audit Loops
Automated checks and guided agent reviews validate runtime health, public boundaries, memory separation, search quality, access-tier enforcement, and mobile client regressions across every separated service area — with a repeatable test battery run before and after each change.
Skill Updates
Reference skills and stack notes are updated as the platform changes so agents can keep context aligned with the current Kaizen release.
Mobile Apps
Kaizen is being built beyond the browser, with mobile clients focused on real conversation workflows.
iOS App
The native iOS app runs on TestFlight with chat, advanced voice, runtime status messaging, diagnostics, and private assistant continuity. Recent builds added reliable history management across devices, conversation pinning, live formatted responses while the model is still typing, and faster, smoother streaming.
Android App
The Android client is planned/in development for the same public and private assistant patterns. Public access will follow announcements or direct approval.
Capability Areas
Public copy is intentionally high-level. Specific model identities and private endpoints are not published.
Search
Current-information retrieval, source-aware response generation, and output cleanup for readable answers.
Voice
Speech input, speech output, runtime status messages, operational call alerts, live real-time phone conversation, and native mobile interaction patterns.
Image Generation
Prompt-to-image creation, edits, background replacement, upscaling, and restoration, with replayable jobs managed from the operator dashboard.
Deep Research
An explicit research lane that sweeps sources local-first, verifies claims in a bounded second pass, and returns cited reports instead of confident guesses.
Communications
SMS, email, Pushover, voice-call alerts, delivery receipts, contact relay, and operational notifications through profile-aware services.
Reminders
Stack-wide scheduled messages for known users, created from natural language and delivered through approved channels.
Memory
Private semantic recall, per-user memory management, strict user separation, and stricter policies for saving new facts — with add, edit, search, and cleanup handled per person from the operator dashboard.
Automation
Private tool routing for home and infrastructure operations, with operator-gated access controls.
Coming Next
Near-term work is focused on safer access, clearer runtime controls, and better mobile continuity.
Agent governance
Access tiers and centralized gateway enforcement are now live. Next is policy polish, clearer public limits, and specialized private-agent packaging for memory, automation, public chat, and future agent routes.
Public queue visibility
Cleaner public messaging for queued requests, daily limits, route availability, and early-access contact paths.
Specialized agents
Additional private agents for infrastructure awareness, media workflows, home operations, and personal productivity will remain gated by role and policy.
Public Safety Rules
KaizenAI Open operates with a smaller trust envelope than Max.
No private memory
Public prompts do not receive personal profile context, relationship memories, historical chats, or user-specific preference injection.
No device control
Public prompts cannot route to private automation tools or operate lights, locks, alarms, network equipment, or other devices.
No internal disclosure
Public responses use Kaizen release terminology and do not disclose base model identities, private prompts, hostnames, ports, keys, or infrastructure maps.
