From a thousand signals, the few that actually cause the outcome.
Correlation is cheap and everywhere. TwinSight isolates the variables that drive a result, quantifies how strongly, and reports how much stays unexplained — so people make data-informed, not data-dictated, decisions.
An illustration: scattered candidate variables on the left, with a few causal-driver paths converging into a single outcome node on the right.
Most systems show what is. The hard part is why — and what if.
TwinSight separates the few variables that cause an outcome from the hundreds that merely correlate — then quantifies how strongly, so a decision rests on mechanism, not coincidence.
Causal inference, end to end.
From hundreds of raw variables to a validated model you can act on — and the personalized interface that makes it understandable.
Causal discovery
Sift hundreds of candidate variables down to the ~15–20 with the largest causal impact on the outcome you care about.
SURD decomposition
Separate redundant, unique, and synergistic causality — and measure how much remains unexplained.
Causal graphs
Model how the key variables connect — direction and structure, not just association.
Validation & quantification
Test the graph against historical data to produce a validated, quantified causal model.
What-if simulation
Drive agent-based simulation with the key variables to preview an intervention before you fund it.
Cognitive personalization
Adapt every interface to how a specific person reads data best — within accessibility standards.
A method, not a guess.
A disciplined, auditable path from raw data to a model you can defend.
Inputs
Hundreds of candidate variables from your existing pipelines. Nothing is rebuilt.
Discover
Narrow to the variables with the largest causal impact — SURD, PCMCI+.
Map
Build the causal graph: structure and direction — DoWhy GCM, NOTEARS.
Validate
Test against historical data for a validated, quantified causal model.
Simulate
Run user-friendly what-if scenarios before committing to a course of action.
Supplementing, not replacing. TwinSight surfaces the causal evidence; the human expert makes the call. Every step is auditable — and we report what the model still can't explain.
SURD: separating the kinds of causality — and admitting what it can't explain.
Synergistic-Unique-Redundant Decomposition pulls apart how variables cause an outcome. Two drivers acting together are not the same as one acting alone — and treating them the same is how analyses mislead.
- RedundantEither variable alone would tell you the same thing.
- UniqueOne variable carries causal weight the others don't.
- SynergisticThe effect only emerges when variables act together.
- UnexplainedCausality the data can't yet account for — reported, not hidden.
Reporting the unexplained share is the honesty most methods skip. It's also what keeps a human expert in the loop.
outcome: microplastic load
Microplastic load is driven mostly by unique contributions — a few variables (tire wear, runoff) each carry weight on their own. About 14% stays unexplained.
Once you know the drivers, test an intervention before you fund it.
The validated model feeds an agent-based simulation. Decision-makers ask plain-language questions and see projected outcomes — no statistics background required.
"If we add stormwater filtration at road curves, would that lessen the impact of tire wear on microplastic pollution?"
- Drivers
- traffic & vehicle counts, water runoff, road geometry, temperature, geospatial location
- Lever
- curbside stormwater filtration at high-curvature road segments
Projected effect
illustrative projection — not measured data
Method
Agent-based simulation · federated · data stays in place
Causal inference is horizontal. The method travels.
Wherever decisions ride on tangled, high-volume data, the same engine separates cause from coincidence.
Statistics, cognition, engineering, sensing.
An interdisciplinary group bringing rigorous, cross-disciplinary expertise to every project.
Matthew A. Sazma, PhD
PhD, UC Davis; MA, University of Chicago. Published research on stress and memory; taught research methods, statistics, and cognitive psychology.
Mark Koranda, PhD
Psycholinguistics and prompt design, with a background spanning military translation, cognitive research, and software engineering.
Alyssa Borders, PhD
15+ years in data analysis and visualization; deep experience with large, HIPAA-compliant government databases at California's HCAI.
Erik W. Ness
Full-stack developer for a billion-dollar auto group. Built the MEVN reporting framework and real-time API feeds used daily across dozens of locations.
Matt L. Miller, PhD
Assistant Professor of Psychology, Oakland University. Expertise in simulation and municipal data-to-action; published on latent growth curve modeling.
Cristin J. Irwin
CEO of Visionex (14 yrs). Machine vision and high-performance optical sensor arrays for defense and industry clients.
Have a decision riding on messy data?
Tell us the outcome you're trying to move. We'll tell you whether causal inference can separate the signal worth acting on — and what it would take.