# AI & Automation — Roman Žydyk

My AI work focuses on connecting models to useful tools and workflows, not training foundation models. Through personal projects and experiments, I explore coding agents, persistent memory, integrations, and automation.

My contribution is the integration: connecting tools, designing workflows, managing context, and validating results. The underlying models and frameworks are existing tools, not my own inventions.

Main profile: https://romanzdk.com/index.md · Czech version: https://romanzdk.com/cs/ai/index.md

## Example: a connected personal assistant

One personal project has explored how an assistant can combine conversation, long-term context, external tools, and recurring tasks. Examples include preparing email drafts, producing source-linked digests, recalling project decisions, and answering voice queries.

It was built from existing open-source parts such as Hermes Agent (https://github.com/NousResearch/hermes-agent) for the assistant and Hindsight (https://github.com/vectorize-io/hindsight) for memory, with Home Assistant for the voice side.

General design, simplified: inputs (chat, voice, schedule) → assistant (model + instructions) → memory and tools (context, APIs, browser) → checked result (review before sensitive actions). The last step is a design principle, not a single technical gate.

## Areas of experience

- Agent-assisted software development: splitting work into implementation, tests and review, with review done by an agent that didn’t write the code; handing context between agents through written briefs and shared rules; a stronger model plans while simpler tasks go to cheaper ones.
- Memory & personalization: extracting facts, recalling context and consolidating memory, while keeping provenance and telling confirmed information apart from guesses and outdated details.
- Tool integration: connecting models to APIs, MCP servers, the web and real browsers, and to productivity and home systems.
- Research & productivity: source-linked digests, product comparisons, keeping track of information, draft replies and follow-up tasks.
- Voice & home automation: experiments with voice interfaces, keeping simple local commands deterministic and sending only harder questions to an AI model.
- Evaluation & reliability: comparing models and agents on concrete tasks by quality, speed, cost and reliability; experiments have included agents that summarize alerts with a likely cause for a person to decide on; checking results, monitoring and restorability.

## Principles & lessons

I prefer limited permissions, human review for sensitive actions, and verification against real systems. For simple or critical operations, deterministic automation is often the better choice.

- AI helps most with messy input: reading, summarizing, finding context, drafting.
- Memory is useful only when it keeps where a fact came from and whether it’s still true.
- An agent saying “done” is a claim, not a result. Tests and checks decide.