2026
Proxis — domain AI agents
Voice and image input in Telegram, with a shared agent core.
Creator and developer
The challenge
Chatbot business logic is often embedded in message handlers and cannot be reused across Telegram, web and mobile.
What I built
- Designed a transport-independent domain core.
- Built Proxis Food with voice and image input.
Results
- Proxis Food works as a Telegram bot with voice and image-based meal entry.
- Created a reusable architecture independent of the communication channel.
Components & connections
Proxis Transport
- Transport adaptersTelegram interface to the domain core.
Proxis Core
- Domain intentsDomain commands and workflows.
Proxis Ai
- ToolsActions available to the domain agent.
- LLM providersText and image interpretation.
Proxis Data
- Profile & historyUser context and history.
- Limits & usageRequest and resource accounting.
- Transport adapters↔Domain intentsnormalized input
- Domain intents↔Toolsexecutes
- Domain intents↔LLM providersinterprets
- Domain intents↔Profile & historycontext
- LLM providers→Limits & usageusage
Engineering takeaway
Context
An independent domain-agent project with a working Proxis Food bot.
- Fast data entry and editing in a conversational interface.
Approach
Separated domain logic from Telegram and tested it through a working bot.
- Used Telegram as an adapter, moving business logic into the domain layer.
- Moved user profiles, intentions, tools, history and limits into the domain layer.
Lessons
- A shared core is easier to evolve around concrete user workflows.
Stack
- Telegram bot
- LLM vision/text
- Domain architecture
- Conversational UX