Technology architecture · tools serve the workflow
The technology stack
The stack is organized by function: acquire evidence, do the work, preserve state, store assets, deliver the product, and observe what actually happened. Any individual AI model can be replaced without losing the business.
Models think. Systems remember. Live surfaces prove.
The redesign
GitHub is the control record, not the binary bucket. R2 is the working binary store. Beehiiv is the publication engine. The browser is an operating surface—not a durable integration.
End-to-end architecture
From outside signal to observed reader
The blue return line matters as much as the green production line. Without received-form evidence, the system knows what it sent—not what the reader saw.
Production flowObservation and learningHuman or promotion gateOpen integration
Preferred order: connector or API for repeated work; browser for configuration and visual proof; manual action only at the real gate.The dotted migration lane is acknowledged but not designed here.
Management view
What each layer owns
A tool is not an owner. Ownership sits with the layer's contract, state, and proof requirement.
| Layer | Primary system | Contract | Failure test |
|---|---|---|---|
| Acquisition | Grok, Gmail, web, APIs | Evidence package with source, date, relevance, and uncertainty. | Can another worker verify the claim without the chat? |
| Orchestration | ChatGPT / Codex | Named job, named inputs, inspectable output, and explicit gate. | Could the work resume tomorrow from files alone? |
| Durable state | GitHub master + Obsidian view | Current routing, versions, protocol, and provenance. | Does NOW.md point to one current artifact? |
| Binary storage | Cloudflare R2 | Stable object and manifest pointer. Drive remains a static archive. | Can the named image or proof be retrieved directly? |
| Delivery | Beehiiv + Cloudflare | Email, site, ops, and reading surfaces as the reader sees them. | Did the delivered product arrive, render, and link correctly? |
| Observation | Gmail, analytics, replies, logs | Real-world evidence tied back to the named run. | Are corrections based on observation rather than model confidence? |
Build now
- Prove Beehiiv sender and received-form QA.
- Keep GitHub and R2 roles sharply separated.
- Preserve one named run package per issue.
- Return live evidence to the next run.
Do not optimize yet
- Multi-agent autonomy before the line repeats.
- Automated editorial promotion.
- A new platform to replace the current spine.
- The legacy routines until the issue factory is stable.