Structured output from messy PDFs with confidence scoring. Anything under threshold goes to a review queue instead of the database.

Traces every agent run, scores it, and diffs behavior across model versions. The tool I reach for before shipping anything.

Agent watches alerts, correlates logs, and opens a PR with a proposed fix. Never merges on its own — video explains that decision.

Hybrid search with reranking across 200k documents. Demo covers the chunking strategy and how I measured retrieval quality honestly.

Planner delegates to search and summarize workers, then a critic rejects weak sources. Walkthrough of why the critic mattered more than the planner.

Reads tickets, pulls account context, drafts a reply or escalates. Video shows the eval harness that caught a 12% hallucination rate before launch.