The archive.

Shipping real software with coding agents, where the value is the judgment layer on top.

A red build tells you something failed. It doesn't tell you the thing that matters — whether a human needs to fix it now, or whether the test will be green next run. Here's a small, dependency-free flake engine that answers that from run history.

When you turn public documents into sales signals with an LLM, the expensive failure isn't the signal you miss — it's the one you invent. Two disciplines fix it: cite every signal to a verbatim source span, and abstain when the document supports nothing.

Turning ‘the community owns its data’ into enforceable, audited code: default-deny access, owner-granted sharing, no super-role, and an append-only audit trail — over a territorial-monitoring dataset.

Scoring forest loss inside a plot boundary is the easy half. The engineering that matters is refusing to certify on bad geodata.

Putting an eval harness and an abstain gate around an LLM analysis agent so it's trustworthy on noisy marketing data

Why an AI that finds missed hospital reimbursement has to cite the chart and abstain — and why that guardrail belongs in deterministic code, not in a prompt.

A worker walks up to a turnstile. Should it open? The safe answer to that question is a small, pure, boring function — and a tamper-evident record of every time you asked.

When the edge state store goes down, your middleware has to choose between letting traffic through and blocking it. That choice is a design decision, not a default — and the right answer flips depending on what the middleware is for.

Building a gap-free, append-only compliance ledger in Elixir/Phoenix — why the obvious approaches all fail, and the one pattern that survives a regulator.

Ranking, dedup, and the senior instinct to refuse — building a camera-roll video picker.

Why a compliance or money classifier keeps a deterministic core — and puts the model behind an abstain gate.

Four parties, one work order, and the single invariant that keeps money and compliance from leaking out of the seams.

Validating LLM-generated assessment content deterministically before it reaches a student

Giving an LLM observability assistant an eval harness and an abstain gate so it never invents a root cause on noise.

Most analytics ships a re-identifiable event log to a server and aggregates there. Here's the inversion — aggregate on-device, and refuse to report any cohort small enough to single someone out.

A patient-journey analytic over a claims graph, taken from a naive plan to a materialized-view rollup — with the one rule that separates an optimization from a bug.

Letting an LLM repair a broken selector is easy. The engineering is in refusing to believe it.

Why a next-shift census model has to ship uncertainty, defer on thin data, and prove itself against a naive baseline before anyone staffs a floor from it.

How to stream an LLM roleplayer over Server-Sent Events, and why the debrief that scores a trainee should abstain instead of inventing a number.

Building a tamper-evident, hash-chained log and the verifier that makes it worth anything — and what real end-to-end voting adds on top.

Normalizing heterogeneous financial data with provenance — and building a gate that refuses to score a thin file.

A small, runnable prototype of a voice product over email — transcribe, LLM reply, synthesize, thread back — and why the most important line of code is the one that refuses to speak.

How we built Beacon — a responsive React SEO & Core Web Vitals dashboard with hand-coded SVG gauges, no chart library, and a theme system that never flashes.

Server-Sent Events are small but fiddly to parse correctly, and the existing Python libraries couple parsing to one HTTP client. sansio-sse is just the state machine — the WHATWG event-stream algorithm, reusable on httpx, requests, aiohttp, or a raw socket.

Every remote call needs retries with jitter, but jitter is randomness, and randomness is where retry code flakes. backofflite keeps the backoff math pure and seedable and makes time.sleep injectable, so you can assert the exact delay schedule and no test ever really sleeps.

Haskell has QuickCheck, Python has Hypothesis, JavaScript has fast-check. PHP never got a good, maintained one — so we built forall: property-based testing with automatic shrinking, for Pest and PHPUnit.

We built one small application — a public message board — idiomatically in Laravel, Symfony, CodeIgniter, CakePHP, Slim, and WordPress, and deployed every one. A live tour of what each framework is actually good at.

A healthcare denial-triage tool built on Django, Celery, and Postgres: webhook ingest, a background LLM classifier with structured output, transient-failure retries, PHI-free logging, and an abstain path that parks uncertain claims for a human.

A production-shaped LLM classifier for mortgage servicing: structured JSON output, idempotent loan-tape import, an append-only audit log, and the abstain path that keeps a confident-wrong answer off a borrower's mortgage.

Building an idiomatic Drupal 10/11 custom module end to end: Migrate API source + process plugins, content type/fields/taxonomy via configuration management, Twig theming, and Drush commands.

JavaScript has picomatch and minimatch; Ruby doesn't. picoglob compiles bash-style glob patterns into reusable Ruby Regexps so you can match arbitrary strings — S3 keys, routes, log lines, branch names — not just files on disk.

A durable invoice intake to provisioning workflow engine in pure Python: an async orchestrator with retries/backoff, a dead-letter queue, structured logging, SLO reporting, and pluggable ERP/ticketing connectors.

A pure-Python reference for enterprise IAM: RBAC + ABAC authorization, federation/SSO claim mapping, non-human identity governance, segregation-of-duties, access certification, and a hash-chained audit log.

A dependency-free Ruby port of fzy's fuzzy-match scoring that returns both a relevance score and the matched character positions for highlighting — exactly what you need to build a command palette, quick-open, or CLI picker.

Building a pixel-perfect, fully responsive marketing landing page in semantic HTML5 and modern CSS3 with ES6 JavaScript (and a jQuery variant) — mobile-first, accessible, cross-browser, zero build step.

The Vercel AI SDK's useChat hook speaks a specific Server-Sent-Events wire format. The official backends are JavaScript and Python — Ruby had nothing. Here is the gem that lets a Rails or Rack app drive useChat from Ruby.

How we built a clean, machine-verified speech corpus across 38 languages from messy public audio — the dedup, the Whisper verification gate, and the edge-served manifest behind LinguaGuessr.

Could you say 'Hey Claude' out loud and have a coding agent start working — fully on-device, without piping your microphone to the cloud all day? Here is the pipeline, the wake word we trained from scratch, and the macOS permission story underneath it.

A practical workflow for finding and killing N+1 queries in a large Rails app, using bullet, prosopite, and strict_loading together rather than one at a time.

Partial indexes, EXPLAIN ANALYZE, composite-index column order, and counter caches — the Postgres indexing patterns that matter once a Rails table gets large.

Russian-doll fragment caching, low-level Rails.cache, cache key versioning, and why you almost never write an explicit expire. The caching patterns that stay correct as the app changes.

Exactly-once is a lie, so design for at-least-once. Idempotency keys, dedup, retries with backoff, poison jobs, and the dead set — the Sidekiq patterns that keep background work correct under failure.

A production Turbo Streams + Stimulus pattern for real-time Rails UI — no React, no API layer, no client-side state to keep in sync.

Locks, backfills, CONCURRENTLY, and dual-writes — how to change a large Postgres schema under a live Rails app without taking it down, and why strong_migrations belongs in every project.

When aggressive denormalization is the right answer.

A Salesforce sync, a read follower, and the audit that reshaped a credentials posture.

A production integration built on AWS SES, a job queue, and compound indexes.

Building an internal AI plugin system for an engineering team.

Building an MCP server so Claude Desktop could edit production data safely.

Cross-team context for AI agents without a real migration.

A reparent-user workflow with a dry-run diff.

Auditing a decade of Rails code in one afternoon.

What we learned migrating a generated admin to React + GraphQL.

RDS IOPS pattern detection with Ruby and a spreadsheet.

Turning a year of access logs into an archival policy.

How we cut an S3 line item ~60% without breaking a single user workflow.

A production server-driven UI across iOS, Android, React, and React Native.