Impact estimates with an evidence grade, checked against high-quality published research. Uncertainty sits next to every effect, so you see when the data can't answer. A phased engagement that shows value early.
Role-based views around your workflows: ranked lists, root causes and recommended actions. Answers grounded in your causal findings, with the reasoning shown and a next step to try or test. Every recommendation can be overridden.
Synthetic data for build and demo, de-identified extracts in operation, and HIPAA- and FERPA-aware handling. Your cloud or ours, on platform-agnostic Kubernetes. Lineage and data-quality checks you can audit; no automated decision affects a person.
Send us your security questionnaireA phased engagement that shows value early and keeps your team involved throughout.
Indicative timeline; the actual schedule depends on data access and scope.
CML+ is organized as a loop of six stages, each removing one barrier that keeps causal AI out of everyday decisions.
“Can we trust causal algorithms on data like ours?”
We generate realistic synthetic data that mirrors the structure of your organization: same tables, same quirks, same selection into treatment. Then we plant intervention effects that differ by subgroup. Because the right answer is known, we can measure which estimators recover it, how well they find who benefits, and where they fail.
“Our data is scattered, inconsistent and incomplete.”
LLM agents analyze your systems, propose mappings and quality checks, and assemble a data lake that serves as a single source of truth (SST). AI and data engineers supervise every step. We add the public data that explains outcomes your systems can't see, from neighbourhood context to labour markets to weather.
“Are our retrospective impact results real?”
We estimate what past interventions actually did, using a portfolio of methods chosen for each question: matched comparison, regression discontinuity, difference-in-differences and doubly robust estimators. The evidence library (Evidence Curator) then lines each estimate up against high-quality published evidence (RCTs and meta-analyses), including null and negative findings.
“Insights take months to arrive.”
Digital twins, built from people like the ones you serve, let you ask "what if we did this, for these people, starting then?" A causal sequence model rolls each twin forward under the intervention and without it, from the same starting point. The result arrives in minutes, next to the published evidence closest to that intervention.
Reach 70%
Change in chronic absenteeism at week 12 vs. no intervention
−3.3 pts
consistent with published RCTs“Analysis never turns into learning.”
Simulation ranks the options; experiments confirm them. We design prospective randomized trials where a clean answer is needed. Where learning while doing is better, we use multi-armed bandits, which shift effort toward what works while they keep learning. Results flow back into the single source of truth and the models.
Multi-armed bandit: allocation by option
1,200 participants allocated. Share of participants assigned to each option; the bandit shifts toward what works while it keeps learning.
“Causal results are too nuanced to act on.”
Effects come with uncertainty, conditions and caveats. Our agents turn them into forward recommendations personalized to each stakeholder: the executive deciding on investment, the analyst checking the method, the front-line professional with a caseload today. They also teach the reasoning as they go.
A network of specialized agents handles onboarding, data transformation, causal modelling, experimentation, validation and MLOps, so the loop keeps learning without a large data team.
Learns your systems, KPIs and decisions, and sets up the project with your team.
Map, clean and join your data, and engineer the features that explain outcomes.
Estimate what each action does, for whom, and run counterfactual simulations.
Run the best options as RCTs or bandits, and feed every result back.
Continuous causal learning: every experiment retrains the models
Test methods on synthetic ground truth, and check every estimate against evidence extracted from published trials and studies.
Deploy, monitor and retrain models, and flag drift before it affects a decision.
Explain each result to each stakeholder in their own terms.
Train people on the product and coach teams through the change.
Agents propose and explain; engineers and decision makers approve.
Read across a row to see one stage in each industry, or down a column for one industry's full loop.
| Stage | K12 Education | Healthcare | Water Utilities | Fintech |
|---|---|---|---|---|
| 1 · Validate | Synthetic districts and campuses with planted intervention effects by grade and root cause | Synthetic ICU and member cohorts with known treatment effects | Synthetic usage and asset histories with planted conservation effects | Synthetic customer journeys with channel-specific lift |
| 2 · Unify | SIS, LMS, MTSS, nurse and transportation data | ICU time series, claims, EHR, wearables | Billing, SCADA, work orders, aquifer and weather data | CRM, lead sources, servicing and settlement data |
| 3 · Verify | Attendance, literacy and mentoring trials; regression discontinuity at checkpoints | Sepsis and critical-care RCTs and meta-analyses | Conservation-program evaluations | Marketing-mix and retention benchmarks |
| 4 · Simulate | Which outreach, for which students, starting when | Treatment policies on patient digital twins | Conservation scenarios; preventive vs. reactive maintenance | Budget shifts across channels; settlement sequences |
| 5 · Experiment | Bandits on intervention menus; student-level RCTs | Virtual and prospective trial cohorts | Phased behavioural interventions | Channel and message tests |
| 6 · Explain | Superintendents, assistant principals, families | Clinicians, care managers, trial teams | City managers, crews (bilingual assistant) | Executives, marketers, settlement specialists |
We'll start with synthetic data shaped like yours, so you can see results before any data leaves your systems.
Book a working session