Category: Uncategorized
-
OWASP’s LLM Top 10, through the lens of iForgetalot
Walking through the OWASP Top 10 for LLM Applications and how each item lands when your agent runs on the phone, dispatches via narrow action tags, and only falls back to a frontier API under memory pressure.
-
Action tags: making an LLM actually do things
Tool use without giving the model a code interpreter. A pattern for safe, debuggable agent actions that ships in production today.
-
The 70% heap rule: when to bail out of local inference
Running a quantized LLM inside a React Native app means the JS heap and the model weights are roommates. A simple memory-aware fallback prevents the crash that ruined our build 49 TestFlight release.
-
Why we put the LLM on the phone, not in the cloud
Local-first inference changes the economics, the latency, and the privacy story of an agentic app. Here’s why we made it the default for iForgetalot — and when we still fall back to Claude.