Research

We study how AI systems can write their own procedures by observing how enterprises actually operate. Procedure synthesis from observed work, with provenance and continuous refinement, is the layer between process discovery and reliable agent execution.

Correction Taxonomies

Expert corrections to AI systems are treated as noise. We study the structure hidden in correction patterns, why models fail, what types of failure exist, and how knowing the correction type predicts which future decisions will need human review.

Context Engineering

Large models are only as good as the context they receive. We study how to select, compress, structure, and sequence operational knowledge at inference time so that context quality, not model size, drives accuracy in enterprise tasks.

Specification Mining

Enterprises spend months documenting how they operate. We study how to observe real work and automatically extract executable specifications from operational patterns, without interviews, workshops, or pre-existing documentation.

Knowledge-Grounded Specialization

Fine-tuning is expensive, brittle, and requires retraining. We study how persistent knowledge graphs can turn general-purpose models into domain experts at inference time, matching or exceeding fine-tuning without touching model weights.

Publications

2026BenchmarksCrosserBench: A Benchmark for Cross-System Entity Resolution
Coming SoonSpec InferenceTurning Expert Corrections into Executable Business Rules
Coming SoonKnowledge GraphsLiving Knowledge Graphs for Enterprise AI: Architecture and Deployment Patterns
Coming SoonAgent SpecializationFrom General-Purpose to Domain Expert: Specialization Without Fine-Tuning
Coming SoonEntity ResolutionOn the Failure Modes of RAG in Enterprise Contexts
Coming SoonKnowledge GraphsThe Correction Flywheel: How Accuracy Compounds from Expert Use

Interested in collaborating on research?

We work with academic and industry partners on enterprise AI knowledge infrastructure.

Get in touch