Conversation State in AI Products: Patterns That Survive
Why stateless LLM calls plus structured turn state beat in-prompt history, and the conversation-state patterns we ship in production AI products.
— /journal
Engineering notes & field stories
Production-tested patterns, product stories from Blink AI, Nexia Academy and Wallnetic, and the trade-offs we hit shipping real software. Written by senior engineers, not marketing.
-
-
Building Internal Tools That Don’t Slow You Down
Why we keep internal admin tools intentionally boring, the patterns that age well, and the temptation to over-engineer them that costs every team.
in Engineering -
Why We Run Postgres for Everything: 2026 Update
JSONB, pgvector, full-text search, queues — Postgres covers more of the stack than most teams expect. Where we draw the line and when we reach for a specialized store.
-
Migrating an Expo App from Bare React Native: When and How
When to migrate an existing bare React Native app to managed Expo, the prerequisites that catch teams off-guard, and the step-by-step playbook we run on real production apps.
in Engineering -
Wallnetic’s Metal Renderer: How We Draw Live Wallpapers on the Desktop
Why Metal over Core Animation for multi-display live wallpapers, the sandbox-compatible rendering pipeline behind Wallnetic, and the power budget we have to fit inside.
-
Multi-Tenant LMS Reporting: Cross-Tenant Aggregation Without Leaks
How we build cross-tenant analytics and per-tenant reporting on the same data layer — with tenant isolation enforced at query time, not by convention.
-
How We Validate cmi5 LMS Compatibility Before Shipping
Three LMS vendors claimed cmi5 support; only one shipped a working launch flow. Our pre-flight checklist and the exact tests we run before any cmi5 content reaches a customer LMS.
-
Designing an LRS from Scratch: Statement Schemas and Query Patterns We Wish We Had Earlier
The statement schema, query patterns and indexing decisions behind a production xAPI Learning Record Store — and the ones we wish we had made at the start of Nexia Academy.
-
Evaluating LLM Coaching Quality: Our Rubric, Our Tools, Our Mistakes
Why LLM-as-judge alone is not enough for a coaching product, the rubric we use across human and automated review, and the evaluation mistakes we wish we had avoided.
-
Per-User Memory in Multi-Tenant AI: Embeddings That Do Not Cross-Pollinate
How we isolate per-user memory in a multi-tenant AI product so context follows the right person across sessions — without ever leaking across tenants or users.