23Articles
Practical write-ups on the technologies I work with: .NET and C#, SQL Server, AWS, observability, and the apps on this site. Newest first.
RSS feedTypeScript 7's compiler is a Go port, and it is fast: type checks five to six times faster on two real projects. What stops the upgrade isn't your code, it's every tool that uses the compiler's API, which 7.0 doesn't ship. Measured on json2types and a game monorepo, with the side-by-side setup that keeps typescript-eslint and compiler-API tests working.
A private, offline weight diary in TypeScript that keeps every weigh-in on your device: IndexedDB with a memory fallback, backups that merge without undoing corrections, a smoothed trend under the daily noise, projections with honest error bands, charts without a chart library, and a Content Security Policy that stops scripts from sending data.
How json2types turns pasted JSON into C# classes and TypeScript interfaces: keeping the text of every number, merging every sample instead of reading the first, sizing integers exactly and checking decimals against their range, typing dates only when every value is one, choosing names every compiler accepts, and testing the output against the real .NET and TypeScript compilers.
This year's npm attacks published malicious packages with valid provenance, from the projects' own release workflows. What provenance does and doesn't prove, and the defences that would have helped, tried on npm 12: lockfiles, a release-age cooldown, blocked install scripts, and trusted publishing scoped to a reviewed release. Plus the same ideas for NuGet.
C# 15 adds union types and closed hierarchies, so a method can return 'an order, or out of stock, or payment declined' and the compiler makes every caller handle each case. Tried on .NET 11 RC2: what replaces the hand-rolled Result types, what the compiler catches, what JSON does with a union, and the caveats: boxing, default values and switch statements.
Polly v8 replaced policies with resilience pipelines, and Microsoft.Extensions.Http.Resilience made a sensible default one line. What each strategy actually does, measured against a deliberately flaky local server: jittered backoff, a breaker that fails in a hundredth of a millisecond, timeouts that cancel the request, and why the order of strategies matters.
EF Core's in-memory provider passes tests that the real database fails: unique indexes, column lengths, foreign keys and decimal precision are all ignored. Testcontainers starts a real PostgreSQL per test run in about three seconds. The same four tests, both ways, with real timings and a setup that keeps a hundred tests under a second.
An agent is a model, a set of tools and a loop. With Microsoft.Extensions.AI that's a few dozen lines of C#. The work is in the guardrails: rules enforced in code, approvals for actions that change things, a cap on the loop, and tests that don't need a model. A support agent built and run against a local model, with its real output.
Sending and receiving a message takes five lines. What decides whether a Service Bus system holds up is what happens when a handler fails, a message is malformed, a send is repeated or events must stay in order. Each behaviour here was run against the official Service Bus emulator, with its real output.
Drift is an arcade space shooter drawn on an HTML canvas that has to hold 60 fps on an iPhone with hundreds of ships, bullets and particles on screen. The four decisions that make it work: a fixed-step loop, pools instead of allocation, glow baked once instead of every frame, and a performance budget enforced by a test.
Why Math.random is banned in my games. Every random draw goes through a seeded generator, so level 37 is the same puzzle on every phone, a daily challenge needs no server, and a whole arcade run can be replayed from its inputs. How seeds, forked streams, fingerprint tests and replays work in practice.
Arrows is one TypeScript codebase that ships as a website and an App Store app. How Capacitor wraps it, one platform interface for both, ads and consent in the right order, safe areas, and a TestFlight upload from GitHub Actions without a Mac on the desk.
More colours don't always make a harder water sort puzzle. Decanta rates every deal by its traps: the share of the moves worth trying that lead to a dead end. How the solver, the trap count and the difficulty bands work, with numbers from the real generator.
Every Moat board has one solution that logic alone reaches, generated on the phone in milliseconds. How: castles first, realms grown around them, border cells moved until every other solution dies, and a human-style solver that rates the difficulty.
A casino game's return to player has to be right to the last decimal. How the Original Table Games prove every declared RTP twice: exactly, by running the real game over every outcome with BigInt fractions, and by a seeded Monte Carlo run sized to a 99.9% confidence.
Arrows is a puzzle game for the iPhone and the web with endless levels, a daily challenge, a league and events, and no server. Every board is generated on the device from a seed. Here's how: partition, peel and measure, pixel-art events and a simulated league.
.NET 8 and .NET 9 both reach end of support on November 10, 2026. Here's a practical checklist for moving a service straight to .NET 10, the next LTS release, and the C# 14 features worth adopting on the way.
Every puzzle in my Sudoku is generated in the browser, with exactly one solution, typically in 5 to 15 milliseconds. Here's how: bit masks for candidates, a most-constrained-first solver, symmetric clue removal and a simple difficulty grader.
Cold starts are the price of running .NET on Lambda. Here's how to measure them, the cheap fixes to try first, and how SnapStart and Native AOT compare on the .NET 10 runtime.
Traces, metrics and logs from an ASP.NET Core API into Grafana, with one Docker container and a few lines of setup. Then the dashboards and queries that answer the questions you'll actually have in production.
SQL Server 2025 adds seven REGEXP functions to T-SQL. Here's how to use them to validate, clean, extract and split text, and what to watch for with indexes, collations and compatibility levels.
The WITH clause, or Common Table Expressions (CTEs), is an underused way to structure complex queries. What CTEs are, how to chain them, recursive CTEs, deleting duplicates through a CTE, and what they really do (and don't do) for performance.
Diagrams make processes, workflows and data structures easier to follow. How to have ChatGPT write them in Mermaid's text syntax, where to render them, prompts that work, and how to fix the syntax errors that come up.
