Why we still pick Postgres in 2026
JSON, full-text search, vector embeddings, time-series, queues. Postgres now does the job of six tools — here’s where it shines and where it absolutely doesn’t.
How we ship LLM features that hold up at 90 days
Most LLM features look great on day three. By day ninety, the prompt has rotted, the costs have tripled, and the team is afraid to touch the eval suite — if there even is one. The playbook we use on every engagement: evals first, retrieval second, prompts last, and a real observability stack from the start.
Read the essay →