D Dosev AI

Personal notes · AI tooling · automation experiments

A quiet workbench for learning and building with AI.

This is where I collect practical notes, small experiments, and reflections from a slow, incremental move into AI-assisted work. The focus is simple: understand what is useful, what is fragile, and what needs human review before it becomes part of real work.

Read the notes View public repo
Incremental, not sudden AI is treated as a gradual extension of existing data, reporting, and automation habits — not a personality change or a dramatic reinvention.
Operations before demos The interesting questions are practical: where tools help, where they break, and how to keep outputs reviewable.
Guardrails matter Useful automation needs boundaries: human checks, clear inputs, safe examples, and honest notes about what is not ready.

Cautious signal

A low-noise record of learning, experiments, and operating patterns.

The site is deliberately modest. It is not a launch manifesto. It is not a claim that everything is solved. It is a place to make the learning path visible at the right level: practical enough to be useful, careful enough to avoid exposing private work, and honest enough to show the gaps.

Learning notes

Short observations from AI-assisted workflows, data work, Office automation, and the everyday friction of turning ideas into reliable process.

incremental learning practical notes lessons learned

Technical workbench

Public-safe experiments with tool calling, Office files, validation, and workflow structure — described at the pattern level, not through private use cases.

Office automation MCP patterns validation

Operating principles

A consistent bias toward clear inputs, reviewable outputs, conservative claims, and separation between private work and public examples.

human review safe examples bounded claims

Automation workbench

Small tools, careful boundaries.

Some notes reference Office automation and MCP-style tooling because they are useful ways to explore agent-to-tool workflows: structured inputs, deterministic tool calls, validation, and reviewable artifacts. The framing stays experimental and bounded: what worked, what failed, and what still needs care.

Open public repo Read related notes

Public boundary

  • Safe to show: public-safe examples, general architecture notes, tool patterns, validation lessons, and learning reflections.
  • Say carefully: Office automation, AI-assisted workflow design, MCP-style tool use, and bounded experiments.
  • Do not overstate: no broad product maturity claims, no unsupported impact metrics, and no implication that exploratory work is production-ready.
  • Keep private: employer-specific use cases, client/vendor details, internal IDs, private screenshots, credentials, and non-public datasets.

Blog

Notes from the build path.

The blog stays in Notion for now and is embedded here as the publishing surface. This page provides a clean home, lightweight search metadata, and a stable place to collect posts. Later, selected posts can be moved into native pages if that becomes useful.

Embedded Notion blog. Open full page if iframe scrolling is constrained on mobile.

Open blog in full page

Principles

The point is not to look louder. The point is to think more clearly.

The useful AI work is rarely dramatic. It is usually a slow sequence of better questions: what should be automated, what should stay human, what evidence is enough, and how to avoid confusing a promising demo with a dependable workflow.

Start from existing work

AI is most useful when connected to real reporting, data, process, and communication problems already present in day-to-day work.

Prefer reviewable systems

A good workflow should leave a trail: inputs, decisions, outputs, exceptions, and places where a person must check the result.

Keep the public layer clean

Public writing should explain the pattern without exposing private context. Synthetic examples and conservative language are the default.