I'm a software engineer with 10 years of commercial experience across e-commerce, energy, and enterprise domains. My core stack is .NET, C#, Azure, and microservices - building distributed systems that handle real load on real infrastructure.
In day-to-day development I mainly use Claude Code and GitHub Copilot, plus local models.
When I build something useful and repeatable, I open-source it. Right now that means AI observability platforms, evaluation benches and tooling for coding agents, and books and flashcards for .NET engineers.
|
An architecture constitution for .NET Aspire on Fly.io and Azure - 15 principles and 17 operational guides, readable as documentation and installable as agent plugins for Claude Code, Copilot and VS Code. |
A spec-first evaluation bench for tool-using agents: 35 YAML scenarios, 313 assertions
over the execution trace, a deterministic layer that gates every pull request, and a
rubric judge that reports |
|
AI coding-session observability - an OTLP collector with a hand-written protobuf decoder, a Postgres store, and a Blazor dashboard that scores session quality rather than token spend, across five assistants. |
Standalone auth microservice for ASP.NET Core: JWT with rotating refresh tokens and reuse detection, RS256 with published JWKS, Google/GitHub OAuth, multi-tenant organizations, and an append-only audit log. |
|
Two LLM-powered chess agents battle through a Streamlit interface with configurable personalities and real-time board visualisation. |
The open source projects I actively maintain, grouped by what they are for.
| Project | Description |
|---|---|
| agent-eval-bench | A spec-first evaluation bench for tool-using agents: 35 YAML scenarios, 313 assertions over the execution trace, a deterministic layer that gates every pull request, and a rubric judge that reports skipped rather than green without credentials |
| listing-search-bench | The same evaluation method applied to a hybrid search ranker: a contract written before the pipeline, 28 scenarios and 128 assertions over the execution trace - 22% of them asserting absence - plus a mutation pass that proves the suite can fail |
| payments-bench | The same bench method applied to an outbound bank-transfer service: idempotent initiation, a payment state machine, unknown outcomes resolved instead of blindly retried, a transactional outbox, a double-entry ledger and reconciliation, against a partner simulator that injects faults. A register says what has and has not been proved, and there is no real bank |
| judge-worker | LLM-as-a-judge post-processing worker: Redis/BullMQ batching, mutation-tested judge reliability, and human-calibrated scoring |
| CopilotScope | AI coding-session observability - an OTLP collector with a hand-written protobuf decoder, a Postgres store, and a Blazor dashboard that scores session quality rather than token spend, across five assistants |
| claude-scope | Session forensics for coding agents - a /session-report skill that turns a session's own transcript into token, timing and conversation-shape analytics with a visual timeline; the script does every calculation and the model never does arithmetic |
| context-pin | A versioned rules manifest for coding agents: publish standards once, pin each repository to a version and a channel, and let CI detect drift between its lock file and what is currently published - without a network call on the hot path |
| spec-driven-template | A GitHub template for starting a product from a specification: decisions, spec and task documents filled in before any code, a supervisor protocol that delegates single tasks to sub-sessions, and two CI checks that keep the documents honest |
| exit-interview-agent | An open-source agent for structured exit interviews - bring your own model through MCP, the CLI with an API key, or a local model - with privacy-preserving employer signals and an eval harness that scores interview quality from OpenTelemetry traces. Not deployed |
| AgentHelm | Web cockpit for AI coding agents - drive GitHub Copilot CLI, Claude Code, and Gemini via the Agent Client Protocol with permission policies, audit trail, and git diff review |
| chess-mas | Multi-agent chess in C#, orchestrated with .NET Aspire and Microsoft Agent Framework, with cost tracking against Azure AI Foundry |
| Project | Description |
|---|---|
| architecture-standards | An architecture constitution for .NET Aspire on Fly.io and Azure - 15 principles and 17 operational guides, readable as documentation and installable as agent plugins for Claude Code, Copilot and VS Code |
| architecture-standards-init-example | What that constitution produces when pointed at an empty repository - one Aspire composition root, a 477-line shared kernel with the ceiling enforced in CI, a Next.js surface reading config per request, three Fly apps and eight workflows, and deliberately no domain model |
| authservice | Standalone auth microservice for ASP.NET Core: JWT with rotating refresh tokens and reuse detection, RS256 with published JWKS, Google/GitHub OAuth with provider-side email verification, multi-tenant organizations, and an append-only audit log |
| gh-runner-kit | Ephemeral self-hosted GitHub Actions runners in Docker: one folder per repository, a local browser control panel, scheduled disk cleanup, and a RUNNER_LABELS variable that switches a workflow between self-hosted and GitHub-hosted - the practice described in the self-hosted-runners guide of architecture-standards |
| Let's go Legacy | Tooling for verifying AI-assisted .NET migrations: Second Key records how the legacy system behaves, replays it against the migrated version and checks every difference against a written contract; Portcullis is a diff-scoped Roslyn gate for CI; Legacy Risk Scan is a free in-browser end-of-life and vulnerability report for a .csproj, packages.config or composer.json |
| school-search | A search and ranking portal for Polish primary schools: it finds the catchment school for an address and shows its results, trend and how much confidence the data supports, with moderated reviews and a read-only MCP server. Aspire, several services, in development |
| parcel-number-generator | A warehousing system modernized from a 2018 .NET Framework solution: a finite parcel-number pool with exclusions and three allocation strategies, a notification service absorbed from a second repository, an operator console, and CI that fails a model drifted from its migrations |
| Project | Description |
|---|---|
| MAF for .NET Engineers | Practitioner's guide to Microsoft Agent Framework 1.0 for .NET - agents, tools, MCP, graph-based workflows, multi-agent orchestration, middleware, and hosting, with samples that run against a local model |
| LangChain, LangGraph and Async Python | Practitioner's guide to LangChain 1.x and LangGraph for engineers arriving from .NET - async Python, cancellation, persistence, interrupts, context engineering, and five specified measurement experiments |
| Mathematics for AI Engineers | From arithmetic to scaled dot-product attention in 47 programmed-learning programs; English and Polish from one source, every number computed by a script and re-checked on every build |
| Python for .NET Engineers | Reading, writing and shipping Python when C# is your first language - English and Polish editions from one source, with a guiding project (trace-assert) that the book's own CI builds and runs |
| Staff AI Engineer Preparation Kit | A book, fill-in templates and five printable cards for an experienced engineer aiming at Staff-level AI engineering roles; company-independent by construction, with build gates that refuse named employers and unsourced figures |
| Chaos from Zero | From school arithmetic to the Lorenz attractor and weather ensembles, in English and Polish; every number the book prints was computed by a Python program the build runs |
| Ops Copilot | The companion project to the LangChain book, built one stage per chapter; every stage's tests pass without a single real model call |
| LetsDoSQL | Twenty SQL exercises on the Wide World Importers database, from joins to window functions, each with an explanation and a solution, compiled into one PDF |
| Angular FlashCards and Transformers FlashCards | Beamer Q&A flashcard decks: Angular and JavaScript fundamentals, and the Transformer architecture from the ground up |
| .NET Aspire notes and Orchestrating PostgreSQL with .NET Aspire | A study guide with a full walkthrough of an Aspire solution (API, worker, PostgreSQL, Redis), and an article on adding an Aspire AppHost to an existing .NET 9 solution |
| ConsoleAppAlgorithmsExamples | A .NET console app with small algorithm exercises and their tests: cycle detection, BFS and DFS route search on a grid, a train composition |
| Aprendo a leer, Aprendo los números, Aprendo economía | Daily notebooks for a child, one A4 page per working day, in Spanish with English and Polish editions: reading, numbers and economics; generated from data, with every activity checked by script |
| german-from-zero-to-hero | German from zero to B2 in 52 weeks, in Polish, for an engineer aiming at a job in Switzerland: 48 units with exercises and keys, grammar and vocabulary flashcards, and Anki decks |
| read-more-books and sleep-12h-per-day | Two short Polish books in progress: how to find time to read twice as much, and how to arrange the day around twelve hours of sleep and baking bread (not medical advice) |
| csharp-flashcards | Beamer Q&A flashcard deck covering C# from fundamentals through cloud and leadership topics - built for interview prep, classes, and self-study |
| agents-and-llms | Searchable knowledge base for the Agents & LLMs shorts series - static, search-first page covering basics, foundations, model internals, and agent tooling, with content driven entirely from JSON |
| dev-insight | Practical coding examples and patterns featured on the Dev Insight channel - C#, .NET, Transformers, SOLID, and more |
| C# Dictionaries - Deep Dive | Deep dive into Dictionary internals - hashing, buckets, collisions, resizing, complexity analysis, and benchmarks |
| Project | Description |
|---|---|
| bayesian-inference | A from-scratch exact Bayesian-network inference engine (Enumeration-Ask, no probabilistic-reasoning library) in C#, wrapped in a Blazor WebAssembly and ASP.NET Core app |
| black-hole-sim | A Schwarzschild black hole raytracer in C#: numerically integrates photon geodesics via RK4 and renders a thin accretion disk, with a console renderer and an ASP.NET Core API |
| TinyTransformer | Minimal educational C# implementation of the Transformer architecture - encoder-decoder, multi-head attention, feed-forward layers, and positional encodings |
| rsna-knee-abnormalities-efficiency-lb | Multimodal 2.5D deep learning pipeline (DINOv2 plus a multilingual text branch) for the RSNA Knee Abnormality Detection Kaggle challenge, with a weak-supervision strategy and diagnostics tooling |
| Project | Description |
|---|---|
| IberiaFamilyCalculator | Interactive Spanish IRPF, autónomos, and payroll calculator - drag sliders, add entities, share the full setup via a single URL |
| SupercompensationApp | The supercompensation curve - fatigue, recovery, overshoot - applied to Agile sprints, as a Blazor WebAssembly app with the model's formulas written out in the README and published live to GitHub Pages |
| combustion-reports | Fuel consumption as a curve rather than the brochure's two numbers: fits consumption against average speed and route length, runs entirely in the browser with no account or server, and carries a section on how not to read its own charts |
This repository also hosts a data-driven portfolio map: forty-three repositories, bilingual,
with a consolidation analysis and an operating model behind them, rendered as a small
static site. See docs/PORTFOLIO-MAP.md for what it is, how it's
maintained and how it deploys.
Languages and runtime
Cloud and delivery
Data and observability
Mostly AI tooling, educational .NET content, and long-lived open source platforms rather than throwaway demos.




