Iteration Layer Shifts from TypeScript to Elixir for Distributed AI Workflows
Check the shift from TypeScript to Elixir and evaluate how BEAM supervision can improve your AI workflow reliability.
Inspect your current content‑processing stack and evaluate if a BEAM‑based architecture could improve reliability; consider prototyping a small workflow in Elixir to test supervision benefits.
Summary
Iteration Layer began as a TypeScript API that accepted a request, called a model or library, and returned a response. The product quickly evolved into complex content‑processing workflows that included document ingestion, OCR, schema extraction, image transformation, document generation, spreadsheet output, retries, confidence checks, and webhook delivery. These multi‑step jobs required a runtime that could handle long‑running, failure‑prone, concurrent tasks without blocking HTTP handlers. The team chose Elixir for its BEAM supervision primitives, lightweight processes, and built‑in fault tolerance.
Elixir’s supervision trees allow each workflow step to restart independently, isolating failures from other jobs. The architecture separates request validation from background job processing, enabling explicit retry logic and bounded memory usage. Agent‑generated code benefits from Elixir’s narrow surface area—pipelines, modules, behaviours, Ecto changesets, and Phoenix contexts—making code easier to review and maintain. CPU‑heavy PDF parsing, image processing, and layout rendering are offloaded to Rust for performance, while the core platform coordinates all steps in a distributed system.
Key changes
- Shifted runtime from TypeScript to Elixir to better handle distributed AI workflows.
- Adopted BEAM supervision trees for fault tolerance and isolated job restarts.
- Separated HTTP request handling from background job processing to reduce latency pressure.
- Implemented explicit retry logic for transient failures without duplicating side effects.
- Leveraged lightweight Elixir processes for concurrent CPU‑heavy, I/O‑heavy, and network‑heavy tasks.
- Simplified agent‑generated code by narrowing language surface to pipelines, modules, behaviours, Ecto changesets, and Phoenix contexts.
- Offloaded CPU‑heavy PDF parsing, image processing, and layout rendering to Rust for performance.
- Designed the platform to support long‑running, failure‑prone, multi‑step jobs with bounded memory.