The research wing of Interaslabs
Data operations you can audit,
with exactly one AI call in the loop.
Control Desk turns cross-system mismatches into managed, graded cases: detect → diagnose → fix → verify. Nine of the ten pipeline steps are deterministic. One calls AI. Every step is attributed.
Latest research
Substrate, patterns, config: an architecture for domain-neutral data operations
How a three-layer split — substrate, patterns, client configuration — produces a domain-neutral core that accepts any domain as plug-in data, with exactly one AI call in a ten-step pipeline and machine-enforced purity throughout.
AI calls scale with problem classes, not data volume
Deterministic clustering groups findings into distinct problem classes before any AI is invoked. At 333,333 entities across three systems — 996,736 records, 9,834 findings — four problem classes require four AI calls: a 2,458× collapse in AI spend relative to data volume.
Grading an AI that cannot see the answers
The evaluation harness plants known breaks and mints a secret answer key; a structural guard makes it impossible for the diagnosis path to read that key. Recall and AI precision are attributed to separate layers so failures trace to the right cause.
Control Desk
A managed case, not a Slack thread.
When the data in system B does not match system A, Control Desk opens a case, runs detection, clusters findings into problem classes, diagnoses each class with a single AI call, and writes every configuration change through an atomic, audited path. The workspace reports exactly what it observed — including when AI fell back.