At a Glance
- The problem
- Routine preparation work takes time, and handing it to an automated tool raises fair questions about access and control.
- What was built
- A private assistant for RSG’s own use that proposes tasks in a chat; nothing in that flow happens until the owner approves it.
- What works today
- Approval-based task proposals, a System Health dashboard, startup routines and backups, and local push-to-talk transcription.
The Problem
Running a small technology business involves a lot of preparation: gathering information from different places, drafting notes, checking on systems, and keeping track of tasks.
Much of that work is repetitive, but handing it to an automated tool raises fair questions. What can it access? What happens when it is wrong? Who decides what actually gets done?
Roman is RSG’s practical test bed for those questions: a place to learn what an AI assistant can reliably do for a small business, and where a person still needs to stay in the loop.
How It Was Built
- Keep the owner in control. Roman proposes tasks in a chat, and nothing in that flow happens until it is approved.
- Run privately. Roman is used through a private Windows and browser setup rather than a public website.
- Treat reliability as a feature: system health checks, startup routines, and backups were built alongside the assistant features.
- Be honest about limits. When a request is too large to process safely, Roman rejects it instead of pretending it succeeded.
What Works Today
Working milestones, as reported by the project owner during development.
Private assistant
A private assistant used through a Windows and browser setup for RSG’s own work.
Task proposals with approval
Roman proposes tasks in a chat conversation; the owner reviews and approves them.
System Health dashboard
A dashboard showing the status of Roman’s own components.
Startup and backups
Startup routines and backups, with a reported test restore from backup.
Push-to-talk transcription
Local push-to-talk transcription that places text in an editable message box, so it can be corrected before sending.
Saved sources with size checks
Saved reference material is checked against context-size limits. Requests that are too large are rejected rather than partially processed.
Optional OpenAI connection
An optional connection to an OpenAI model can be enabled for language tasks.
Evidence & Limitations
Where these claims come from
- The capabilities above are milestones reported by the project owner during development. They have not been independently tested, audited, or certified.
- No measured results, such as time saved, accuracy, or reliability figures, have been published yet.
Current limitations
- Internal only. Roman is for RSG’s own use; it is not a public chatbot or a product for sale, and there is no public demo.
- In active development, so features change as the project evolves.
- Like any AI-assisted tool, its output can be wrong, which is why important actions require approval.
Screenshots & Demo
Roman works with private business information, so its interface is not shown publicly. Sanitized screenshots may be added once they are reviewed.
Next Steps
Planned directions, not finished features:
- Richer workflows for tracking prospects and projects.
- Help preparing proposals and a daily brief for review.
- An optional Claude integration aimed at deeper research and capturing reviewed knowledge.
- Possibly, a smaller customer-facing version of Roman.