Startup application readiness · Private Alpha
What is real today, what is being validated, and what a startup cloud program would enable.
This page is written for startup-program reviewers and for anyone deciding whether to follow the project. Every claim on it is tied to a status class, and nothing here is aspirational unless it says so.
- Stage
- Private Alpha · Bootstrapped · pre-revenue
- Running today
- This public site and an owner-only control plane in bounded validation
- Cloud footprint
- Cloudflare Pages and a serverless database. No other infrastructure is deployed.
- Commercial status
- No product release, no customers, no revenue, no outside funding
§ What it is
An owner-governed control and memory layer across replaceable models and tools.
AI OfficeOS is not another assistant. It is the durable layer that sits between a business owner and whatever AI models, agents and tools happen to be good this quarter. It holds company state, keeps decisions together with the evidence they were made on, packages work so it can be handed off or recovered, and keeps the owner as the root of authority.
Models and tools plug in underneath and can be swapped. The company's memory, decisions and acceptance records do not leave when a provider changes its prices, its rules, or its behaviour.
-
Owner-governed
Purpose, spending, credentials, anything irreversible and final acceptance stay with the owner. Nothing in the system can widen its own permissions.
-
Durable control and memory
Company state, typed work packets, decisions and a tamper-evident audit trail live in the control plane, not in a chat transcript.
-
Replaceable workers
Frontier models, open-weight models and ordinary tools are leased capability with an exit path. None of them becomes the company.
-
Evidence in, outcomes out
Work returns as evidence that something else admits or refuses. Results are counted in business outcomes, not in tasks closed.
§ Status
Four status classes. Every claim carries one.
Working now means it runs today and is checked on every change. In active validation means it is built and being tested under a bounded authorization; it is not released. Planned means designed or in review, not deployed. Not yet available means nobody can use it today and no date is promised.
| Capability | Class | Evidence |
|---|---|---|
| Public site on Cloudflare Pages | Working now | This site. Static, no scripts, strict security headers, deployed on every merge. |
| Continuous checks on every change | Working now | Structure, disclosure and secret scans plus an automated test suite gate every push and pull request. Each change is reviewed independently before merge. |
| Owner-only control plane (Control Plane V0) | In active validation | Owner sign-in, company state in a serverless database, typed work packets, a tamper-evident audit chain and de-duplicated event intake. Authorized as a bounded validation; broader release is a separate owner gate. |
| Evidence-bound review of AI worker output | In active validation | Worker results are admitted only when they bind exactly to the work they were asked to do, with provenance recorded separately from content. |
| Website-initiated worker wake-up | In active validation | The bounded wake adapter is deployed, and a one-attempt live wake proof has passed. Protected task read and result return remain unproven; this is not continuous unattended operation. |
| Proactive oversight worker | Planned | One worker that notices stalled, missing or conflicting work and routes it. Designed; not running. |
| Persistent control-plane node | Planned | A small always-on node for scheduled checks and longer-running coordination. Planning target only; nothing is provisioned. |
| Open-weight worker on a GPU node | Planned | A 27-billion-parameter-class quantized open-weight worker for bulk mechanical work. Under evaluation; no model selected, no node provisioned. |
| Managed deployments for outside businesses | Not yet available | Follows internal validation. No customers, no onboarding, no pricing. |
| Standardized subscription edition | Not yet available | Only if managed deployments show repeatable value. No date. |
No public product release. First business scope is an owner decision and is not selected or announced.
§ Demo & evidence
What can be shown, and what is still being sanitized.
The control plane is owner-only and holds real operating records, so nothing from it is published until a sanitized copy exists. Rather than show a mock-up, each item below says exactly what it is.
-
Public site and its checks
Available This page, the overview page, and the automated checks that gate every change.
-
Owner console walkthrough
PENDING_PUBLIC_SANITIZED_ASSET A recorded walkthrough of the owner console on synthetic data. Not yet produced; no substitute is shown.
-
Work packet lifecycle
PENDING_PUBLIC_SANITIZED_ASSET One synthetic work packet from intake to acceptance, with its evidence and audit records. Not yet exported.
-
Audit chain excerpt
PENDING_PUBLIC_SANITIZED_ASSET A sanitized excerpt of the tamper-evident audit chain with verification. Not yet exported.
§ Architecture
Purpose goes down. Evidence comes back. Infrastructure stays replaceable.
- 1Owner and business purpose. Goals, budgets, permissions and final acceptance.
- 2OfficeOS control and company state. Typed work packets, decisions, evidence, audit, handoff and recovery.
- 3Replaceable AI workers and tools. Frontier models for hard problems, open-weight workers for bulk work, ordinary tools for everything else. Any of them can be swapped.
- 4Evidence, outcome and learning. Results return as evidence, are counted as business outcomes, and corrections become reusable cases.
Infrastructure, by status
- Working now Edge site and serverless database
- Cloudflare Pages serves this site; the control plane runs as serverless functions with a serverless SQL database. Free-tier footprint. This is the whole deployed estate today.
- Planned Persistent control-plane node
- One small always-on node for scheduled checks, longer-running coordination and the oversight worker's clock. Not provisioned.
- Planned GPU execution node
- One node for a quantized open-weight worker doing bulk mechanical work at low concurrency. Not provisioned; no model selected.
Internal architecture, endpoints and security mechanisms are not published.
§ Planning targets
What the next two nodes would look like. These are plans, not deployed resources.
- Control-plane node Planned
-
- about 4 vCPU
- 15–16 GB RAM
- 100–160 GB persistent storage
- always on; scheduled checks, coordination, backups
- GPU execution node Planned
-
- about 32–48 GB VRAM
- 27-billion-parameter-class quantized open-weight workloads
- low concurrency, long-running mechanical work
- can be stopped when idle; nothing depends on it being up
Both are sizing targets for a startup-program application and for the project's own planning. Neither node exists, neither has been purchased, and the project runs entirely on a free edge tier today. The targets will be revised against measured workloads before anything is provisioned.
§ Why startup cloud support helps
Five reasons, each with something measurable.
-
Move oversight off the founder's laptop
A persistent node lets scheduled checks and the oversight worker run without a personal computer being awake. Measured as: checks completed on schedule, and owner interruptions per week.
-
Prove open-weight workers on bulk work
A GPU node makes the open-weight evaluation real instead of theoretical. Measured as: cost per completed task versus frontier models on the same tasks, at equal acceptance rate.
-
Keep the company independent of any one provider
Self-hosted capacity gives every dependency a fallback. Measured as: workloads with a tested exit path, and recovery time after a simulated provider loss.
-
Run the first real workload end to end
Validation needs sustained compute for a narrow business loop, not a demo. Measured as: one workload completed end to end with evidence, and its owner-time cost.
-
Learn the operating cost before selling anything
Commercial editions are only justified once the per-business operating cost is known. Measured as: monthly infrastructure cost per validated workload, published internally per cycle.
§ What commercialization means here
Internal operating validation first. Commercialize only after repeatable value.
The owner's decision is that AI OfficeOS is validated on the founder's own operating work before it is offered to anyone. Commercialization begins only after that value is shown to repeat, and then in this order:
- 1Managed deployments. A small number of outside businesses run on instances operated by the project, with the owner-governance model intact. Value and operating cost are measured per business.
- 2Standardized subscription editions, later and only when justified. If managed deployments show that the same configuration serves several businesses well, a standardized edition may follow. It is not promised.
Nothing is for sale today. No pricing, no waiting list, no pre-orders.
§ Roadmap
Seven stages. No dates. Each stage ends on evidence, not on a calendar.
-
Internal MVP In active validation
The owner-only control plane holds real company state and evidence for the founder's own work. Exit: the owner runs day-to-day decisions through it, with the audit chain verifying.
-
Reliable worker wake-up and evidence intake Planned
A worker can be woken from the control plane, fetch its task, and return a result that is admitted on evidence. Exit: one complete cycle proven on the project's own account, twice, on separate occasions.
-
Proactive oversight as a standing duty Planned
One oversight worker notices stalled, missing or conflicting work and routes it to the right place without the owner asking. Exit: synthetic anomalies caught, no false alarms on normal operation, and a heartbeat that fails loudly.
-
Persistent node and open-weight evaluation Planned
Scheduled checks move to an always-on node; an open-weight worker takes bulk mechanical work under measurement. Exit: cost and acceptance rate compared against frontier models on the same tasks.
-
Repeatable value on a second internal workload Planned
The narrow business loop runs a second time on a different workload with the same mechanisms. Exit: owner time, cost and error rate measured against the first run.
-
First managed deployments Not yet available
A small number of outside businesses, operated by the project. Exit: value and operating cost measured per business; governance model unchanged.
-
Possible standardized subscription edition Not yet available
Only if the previous stage shows the same configuration serves several businesses well. Exit: an owner decision on the evidence, which may be no.
§ Contact
Daren Yang
Los Angeles, California. AI OfficeOS is a solo, self-funded project.
A contact address will be published alongside the project's permanent domain.