Chaining GLiNER, DeBERTa and rules to block clinical conflicts in 0.79s

MaziyarPanahi · x · 2026-09-21

The author turned a joke into a working clinical workflow with three model families doing distinct jobs: locally-run GLiNER extracts entities (flags metformin), DeBERTa detects contradictions, and Jev (via typesafeai) returns a blockconflict decision; deterministic code handles the final fail-closed block/review rule, stopping the update and routing it to human review. The recorded run used 709 input + 52 output tokens and finished in 0.79s.

Related event: Three-Model Pipeline Catches Drug Interactions in 0.79 Seconds(2 posts)→

Original post →

More from coding & agent

coding & agent channel →