How we built a production LLM agent in Elixir with closure-scoped authorization, recursive tool execution, and circuit breaking — in about 200 lines of purposeful code.

How we built a data-driven alerting system using Fsmx state machines, two-step confirmation to reduce false positives, and cooldown enforcement to prevent alert spam.

How a single struct as the first parameter to every context function eliminates an entire category of multi-tenant data leaks in Phoenix.

How we built a pipeline that generates LLM prompts and JSON schemas from database records, so adding a new observation type requires zero code changes.

How we solved connection pool exhaustion in an Oban + LLM pipeline by splitting transactions into three phases, separating repos, and working around PgBouncer limitations on Fly.io.

How we built a camera capture and upload pipeline on Raspberry Pi using Elixir Nerves, with SQLite-backed queuing, exponential backoff, and supervision patterns that handle failure gracefully.

How consolidating three codebases into a single Phoenix app with 55 LiveViews cut feature delivery time, simplified testing, and eliminated an entire frontend.

Using dHash and Hamming distance to detect near-duplicate images before sending them to vision LLMs, dramatically reducing API costs without sacrificing quality.

Using PostgreSQL range types and exclusion constraints to prevent double-bookings at the database level, with practical examples in Elixir and Ecto.