AI Job Dispatcher

Challenge

A video production team needed to run LLM jobs (transcript cleanup, script review, reasoning passes) without non-technical teammates touching a terminal or installing anything. Constraints: no admin rights, managed enterprise hardware, restricted egress. It had to run unsupervised and fail loudly; a silent bad answer is worse than a crash.

Process

The v1 interface was a shared folder: drop a file in, get a result out. Underneath ran a deliberately stdlib-only Python supervisor with a durable filesystem queue and a SQLite ledger of every job’s state. Two isolated worker lanes: local media processing, and headless reasoning through the Claude Code CLI (claude -p) in a fresh subprocess per job, so no context leaks between jobs. After a crash, any job left PROCESSING returns to queued on restart. A 7-scenario sandbox harness tests that recovery and the queue’s other failure modes before any change ships.

The key finding was a silent failure in headless LLM orchestration: the CLI could exit 0 with API error text as its output, which a naive supervisor would deliver as an answer. My envelope-error detection turns these into legible failed jobs, locked in by a regression test. Each job carries a provenance record (SHA-256 hashes of inputs and outputs, a prompt hash, the model ID) tracing any output to exactly what produced it.

Result

v1 ran in production for the team, driven entirely by files dropped in a folder. It has since been retired and upgraded into ID Assistant: a Microsoft Teams front end whose connected tools hide the folders from users, with Claude Code authoring the agent and Copilot Studio running it live (see the AI engineering entry on the resume). It was internal tooling for one video team, not a commercial product. Its lasting value was durable infrastructure under a simple interface: a queue that survived crashes, isolated its workers, and never passed off an error as an answer. ID Assistant inherited that explicit-failure rule.