Lumis research
How can intelligent systems understand and safely adapt complex operational environments? Our research begins with that question, and with experiments that can prove us wrong.
Research is part of the system
Our research connects guarded recovery, competing hypotheses, active evidence acquisition, and verification.
Read about our first paperReference architecture available as an arXiv preprint (2608.01955). Peer review is pending; no acceptance is claimed.
Publications
Eshun, Murage, Muoki, Chen, Adjignon, Staar, Angélil. The reference architecture Lumis grew from. Peer review is pending.
Fifteen injected failures on the GridCast reference estate, six investigators, one model. Methods, results, and what went wrong.
Do explicit competing hypotheses and targeted evidence improve diagnosis compared with rules, a single model call, and an unguided tool agent? First results on GridCast say yes, with one model and small samples.
First resultsInvestigating which dependency, lineage, change, and policy relationships materially improve the quality of operational reasoning.
Research directionA synthetic electricity-forecasting estate on Kubernetes with fifteen reproducible failures, injected through the channels they would use in real life. It is a testbed for evaluation, not a live utility deployment.
Reference estate · first resultsStudying how much grounded diagnosis can run locally under limits on memory, compute, energy, and connectivity.
Long-term research
Why it matters
Control rooms made system state legible to people. Our research asks how software can do the same for systems too complex to read at a glance — and stay accountable while it does.
What we measure
Is the supported explanation the right one?
Was the true cause kept open long enough?
How many queries before the decision changed?
How often is certainty unearned?
Later, once Lumis can propose actions.
Later, once there is recovery to verify.
Every result is compared against what an engineer would actually try: deterministic rules, a model given only the alert, a single model call with curated evidence, and a tool-using model without Lumis’ structure.
The first comparison is published, with methods, transcripts and mistakes: one synthetic estate, one model, two runs per scenario. It is a starting point, not a benchmark.
Read the GridCast studyDiscuss research collaboration