Category
16 posts
7 posts
Inside the AI agent backend: LangGraph runtimes cached on a 16-part key, progressive skill disclosure with a 128-tool cap, compare-and-swap settings mirrors, and a feedback pipeline that groups meetings by red flags.
Clean-code and clean-architecture standards meeting real code: the boolean flag that became two functions, the 70 duplicated lines that stayed duplicated on purpose, and import invariants enforced by AST tests.
How the team scaled LLM-assisted development with specialized agents, an adversarial test-triage workflow, and hooks executed by the harness — the layer that instructions cannot override.
A .NET service that ran out of memory in production: per-meeting state that outlived the meetings, a GC ceiling fixed at process start, and why raising the limit was the right stopgap — done deliberately.
Deploying the voice agent on Kubernetes: why one missing webhook secret took a whole namespace down three times, the split-horizon DNS that silently killed observability, and the isolation patterns that came out of it.
Building a voice agent that takes real phone calls: HMAC-verified webhooks, a sideband transcript pump, delegation-based tool dispatch, and a refusal layer protected by cryptographic digests.
A centralized settings module as the last line of defense: boot gates that collect every missing variable, fallbacks that log at ERROR instead of degrading quietly, and why point-of-use beats boot-time for some checks.
9 posts
Inside the notification scheduler: a ten-second tick over five sweeps, five-minute compare-and-swap leases, a recurrence engine with strict and non-strict personalities, and retry accounting that never double-charges.
Designing the push-notification service: idempotency keys as mandatory input, durable in-app records before OS push, fail-open preferences, receipt reconciliation, and WebSocket auth via subprotocol.
An in-process cron scheduler with a generation token, a promise lock, and soft-delete reconciliation — plus the drift panel that turns destructive sync into a recoverable operation.
Designing an Express + Better Auth service where deactivated users cannot sign in even when the database is down, tokens die on role change, and audit logging can never block a login.
A Diátaxis taxonomy, ADRs that running code reads, evidence-first writing, and a nightly production-to-staging database refresh — how one team made documentation cheaper to keep than to skip.
Building a FastAPI service that parses a large polyglot repo with tree-sitter, indexes it with embeddings and FAISS, generates architecture diagrams on demand, and answers natural-language questions about the codebase.
A look inside the three most complex feature areas of the Meridian web app — the assistant console, survey administration, and the audit-heavy admin surface — and the shared architecture underneath all of them.
How the Meridian web app picked React 19, TanStack Router and Query, Jotai, and Tailwind v4 — and why the most valuable piece of the design system was a single prop called containerClassName.
The first post in my Techstack Challenges series, about a client engagement where the coding standards were not a wiki page but executable files in the repo, loaded by both humans and coding agents before every change.