BatchDAG: LLM-Planned Execution Graphs for Scalable Ad-Hoc Analysis Over Enterprise Data

Short summary

BatchDAG is an orchestration layer where an LLM generates a typed DAG of operations (SQL, semantic search, transforms, fan-outs) executed by a deterministic engine with topological-wave parallelism. Entity-aware batching reduces LLM calls by up to 47x. In production at Brevian.ai it processes queries over 50,000+ meetings in under 60 seconds at $0.02-$0.24 per query, achieving quality comparable to expert-designed pipelines with superior provenance (77% transcript evidence rate).

• •LLM generates typed DAGs executed by a deterministic engine with parallel evaluation
• •Entity-aware batching cuts LLM calls by up to 47x; structured JSON intermediates reduce hallucinations by 27%
• •Production deployment processes 50,000+ meetings in under 60 seconds at $0.02-$0.24 per query

Generated with AI, which can make mistakes.