Portfolio · Русский · Repositories
I build backend systems and the tools around them, from product APIs and data pipelines to agent-facing workflows that keep requirements traceable. Most of my commercial work is in Python, but I choose the stack around the problem. AI agents help with analysis and implementation; design decisions, verification, and the result remain my responsibility.
Project requirements rarely arrive as clean tickets. They emerge from project channels, documents, screenshots, feedback, and later corrections. Mnemo keeps that raw material together with its source, date, attribution, and links to requirements, decisions, and open questions. A 20-rule linter checks the archive; a separate self-check compares the published standard with its implementation.
Python · Claude Code plugin · AGPL-3.0 / CC BY-SA 4.0
Herald handles the live edge of the same workflow. It captures incoming messages and files into a local buffer for later import into mnemo, and sends agent-prepared updates through configured communication routes. It moves material between people and agents without becoming a second archive.
Python · MCP · 111 tests · AGPL-3.0
A self-hosted .NET media relay outside the agent-tooling stack. It combines a local Bot API server, modular downloaders, and a persistent queue; files up to 2 GB are supported, and 34 unit tests run in CI. This project is the clearest public example of my C#/.NET work.
C# · .NET 10 · Docker · AGPL-3.0
flowchart LR
people["People and project channels"] -->|capture| herald["herald<br/>communication boundary"]
herald -->|import| mnemo["mnemo<br/>raw context and provenance"]
mnemo -->|facts, requirements, questions| ephemeris["ephemeris<br/>daily state and handoff"]
ephemeris -->|source links| session["Next agent session"]
session -->|prepared update| herald
Each tool has a narrow role. Mnemo keeps durable evidence, Herald handles communication, and Ephemeris records the state of the day in GitHub issues. A handoff passes addresses to source material instead of rewriting the same context for the next session.
I use versioned formats, CI, automated tests, manual critical-path checks, and small reviewable changes. Public repositories also document known limits and discarded approaches. AI-generated code is a draft until I have checked its logic, failure modes, and behaviour in the system around it.
The site is the visual companion to this profile: project pages, screenshots, downloads, and the longer stories that do not fit in a repository card. It currently covers my earlier desktop projects; the next revision will add the engineering case studies behind the work above.