Our Embedding Benchmark Needed a Better Test
Strong retrieval scores became harder to interpret after checking the candidate pool, duplicate sessions, and unequal input windows.
Read articleWriting
Notes from building.
Essays on the product decisions, infrastructure problems, and shipping habits I keep running into.
I am building in public, and documenting pitfalls as AI writes much of the code under close human direction.
Compacting an Idle Claude Session Before Its Cache Goes Cold
Building claude-cache meant choosing when to summarize, preserving the handoff, and refusing to compact after the useful window had passed.
Read articleA Wedding Invite Should Still Feel Like the Wedding at Sign-In
Keeping a wedding's identity through sign-in meant resolving its design from a verified destination and preserving centralized authentication.
Read articleMy First Security Report on Lumberroom Cloud
How a security report led to clearer OAuth destinations, stronger consent warnings, and a closer look at memory access.
Read articleI Made Four Websites Agent-Readable. Some of the Checks Were Theatre.
A scanner moved cybersapien.xyz from Bot-Aware to Agent-Readable in one day. Sivella shipped the same real infrastructure. Vowframes and Chatterbox still aren't agent-optimized. A speculative DNS record, and an entire top tier, were worth refusing rather than faking.
Read articleWhy I Made Six AI Reviewers Disagree
A production crawler design looked ready until six isolated review agents found eight blockers and three disagreements worth resolving.
Read articleRepetition Is Not Knowledge
Why a crawler for Indian SMB websites must separate repeated location pages, dated posts, and useful business facts.
Read articleThe Handoff Is Part of the System
What three Chatterbox codebases taught me about contracts, state, and agent coordination.
Read articleWhy Chatterbox Refuses to Guess
A small-business chatbot protects trust by answering from supplied knowledge, admitting gaps, and creating a path to the owner.
Read articleA Mockup Is a Product Promise
Three weeks of Sivella work showed me how marketing claims and useful-looking UI become an accidental roadmap.
Read articleWhat Is Chatterbox?
How Chatterbox connects a website widget, business knowledge, lead capture, agency controls, and tenant isolation.
Read articleWhat Is Sivella?
How Sivella connects household tasks, money, calendars, meals, decisions, and memory in one working system.
Read articleWhat Is Vowframes?
How Vowframes combines wedding sites, invitation rules, RSVPs, registries, galleries, and planning.
Read articleOne RSVP for a Multi-Event Wedding
Why Vowframes replaced separate event forms with one invitation-aware RSVP flow for guests and families.
Read articleObservability Is Where the Product Stops Guessing
Instrumentation helps a small team see what production is doing before users report the damage.
Read articleA Task Count Cannot Tell You Who Is Free
How Sivella combines household workload with private calendar availability before suggesting a task owner or time.
Read articleA Roadmap Is a Working Surface
A roadmap works best as shared operational state the team can change as work moves.
Read articleHardcoded Model Names Are Product Debt
A stale model name can turn provider settings into product behavior.
Read articleProduction Config Should Not Be a Memory Test
A missing cache URL exposed a split between operator configuration, container topology, and application validation.
Read articleSignup Is a Production Surface
Disposable email checks and confirmation sends decide whether a small signup form becomes abuse infrastructure.
Read articleA Template Needs Its Own Grammar
Building Vowframes’ fifth wedding template exposed the difference between a new palette and a new design system.
Read articleContext Is the Product
Tasks, calendars, and budgets are separate views of the same household decision.
Read articleThe Hardest Software I Want to Build Is a Home That Remembers
Why household software should remember the details one person usually carries.
Read articleThe One More Bug
How polishing turns into launch avoidance, and why real feedback has to start earlier.
Read article