Archive search
73 field notes, newest first
How Fullscript's IT team builds internal apps to replace SaaS tools, and the foundation that makes them safe to run.
See how we migrated 105 merchandising JSON files to a Payload CMS with zero downtime, no app rewrite, and AI agents doing the heavy lifting.
See how Fullscript uses an LLM, a priority-ordered decision tree, and a daily CronJob to categorize every merge request.
See how we paired an LLM with deterministic code to auto-crop Supplement Facts labels, and how we decide what to ask AI versus code.
See how Fullscript fixed search with learning to rank in Ruby: NDCG, judgment lists from clicks, XGBoost reranking, and our upcoming ltr gem.
See how Fullscript built EKG, a single MCP gateway that connects any AI client to our services, with guardrails in one place.
See how Fullscript's SOC built automated L1 triage with 13 AI agents that assess every security alert in under five minutes.
See how Fullscript turned promotions into database records with date windows, cutting sales from multi-deploy projects to one reviewable data change.
How we guarantee one writer per conversation as Fullscript Assist scales across nodes: a durable lease for ownership, an epoch fence for writes.
How Fullscript's Project Glycerin triages and fixes security vulnerabilities in minutes, with engineers owning the standards the AI applies.
See how we built a benchmark from our own git history to compare AI coding models, skill packs, and agent harnesses on our real codebase.
See how Fullscript added 50/50 experiments to Flipper feature flags using a second salted hash and exposure tracking that waits for first use.
See how we treat the context window as a budget in Fullscript Assist: on-demand skills, isolated sub-agents, and auditable compaction.
See how Fullscript added durable logs, threaded Slack alerts, metrics, and self-cleaning governance around Shopify's maintenance_tasks gem.
How Fullscript IT built hospital-grade network redundancy for automated warehouses — see how we design for the failure case first.
How Fullscript rebuilt its clinical AI assistant with Elixir, BEAM processes, and MCP to support 120K+ providers with stateful conversations.
Part 2: See how Fullscript layered live progress, shared lifecycle hooks, and queue isolation onto a gem they don't own — no fork required.
How Fullscript rebuilt the maintenance_tasks UI as a React app over a JSON API — typed params, masked secrets, live status.
End-to-end tests are usually expensive to maintain or too flaky to trust. We treated an LLM as a compiler to get tests that are neither.
Why green tests aren't enough for AI systems. Learn how evals catch behavioral regressions that unit tests miss.
