Zhelezyaka is a concept for a personal AI assistant that runs mostly on the user's own computer, keeps memory between conversations and helps with tasks, projects and devices. The name is the Russian word for a "piece of iron", and the project is my own take on the idea behind JARVIS, made as something that can be built step by step.
The goal is a system that does not only answer questions but understands context, remembers the history of decisions, uses tools and slowly becomes more useful to its owner. The internet and cloud models can extend it, but the basic work stays available locally. The early version, Zhelezyaka Core v0.3.0, had local chat through Ollama, session history, intent routing and structured memory. Autonomy, a reliable permission system, integrations and specialized agents are the target architecture, not finished features.
Zhelezyaka shares the general idea of JARVIS from the Marvel films: a personal AI with voice, memory and character. The difference is that JARVIS is fiction, while Zhelezyaka is a real project that can be built gradually. Both follow the same chain: user, perception, AI core, memory, tools, action. Zhelezyaka adds an explicit permission layer so the assistant knows when to act and when to ask first.
| Area | Zhelezyaka (goal) | JARVIS (films) |
|---|---|---|
| Conversation | Text and voice, understands context and intent | Natural real-time speech |
| Memory | User preferences, projects, decisions, mistakes | Knows Tony Stark's habits and a huge history of work |
| Personality | Own character and independent judgement | Recognizable humour and initiative |
| Work with projects | Code, docs, architecture, tests | Designing advanced technology and simulations |
| Devices | Through permitted tools and integrations | Lab, home, suits, many systems at once |
| Computing | Mostly local models, cloud when allowed | Fictional, almost unlimited infrastructure |
| Stage | Early core, long road ahead | A finished fictional AI |
The target architecture has seven parts. Interfaces and perception (chat, voice, panel). A cognitive core that holds personality and rules, reads the intent, pulls the right memory, and picks a model or tool. An orchestrator with specialized agents: Coder, Archive, Vega (analysis and research), Master (system and tools), Ethics and security, and Cloud. A memory of work, not just chat: what was wanted, what options were considered, what was chosen and why, what failed. A model router that chooses between local and external models. "Gvozd" (the Nail), a security layer that decides which tools are available and which actions need confirmation, with an action log. And a gateway to tools and integrations, plus a planner that also watches CPU, GPU and memory so the AI does not get in the way of the computer. The diagrams are conceptual illustrations of this design.
Send a direct message about this project and it will be delivered to my email.