About — the practice in one read. Positioning, credentials, and the architecture you would be authorizing.
Read this first if you want the argument before the case studies.
This page is the curated portfolio. If you’re evaluating whether the methodology applies to your domain, the relevant tier is whichever your problem rhymes with most closely — not necessarily the most productized one.
For the underlying causal-AI argument and the architecture you would be authorizing, see About. For how engagements begin, see Scope.
Engagements
Selected engagements over the past three decades, anonymized to the industry level. Each line names what the work built, what shipped, and what survived after the engagement closed. Specific firms available on request.
A major US investment bank (financial services)
Private banking platform redesign and cross-border trading systems across New York and London. The design challenge: systems that surfaced the drivers behind portfolio performance, not just reported outcomes. The frameworks shipped into production and shaped how high-net-worth portfolio analytics were structured for the next decade.
A major California electric utility (energy & utilities)
Early-warning systems tracing single upstream events — a regulatory threshold crossed, a supplier disruption, equipment degradation — through their financial, operational, and safety consequences downstream. Single-domain models could not represent those propagation paths; a Bayesian network over them could. The framework survived the engagement and was adopted into standard ops-risk practice.
A leading semiconductor manufacturer (cybersecurity)
Supplemented existing ML threat scores with causal risk models aligned to FAIR (Factor Analysis of Information Risk). The existing scores could rank threats; the added models answered what boards, insurers, and regulators actually ask — not what’s the score but what’s the expected loss reduction if we deploy this control, and what would have happened if we hadn’t? Moved the risk function from description to counterfactual reasoning.
A global automotive manufacturer (operations & strategy)
Technology and business transformation programs connecting operational systems to management decision processes. The work: making the causal links between operating-system signals and management-level outcomes explicit and queryable. Programs spanned multiple business units and several years.
A large urban public hospital (healthcare operations)
Designed the city’s disaster recovery program around cascading-failure reasoning instead of sequential checklists. Standard ISO 22301 frameworks catalog what to restore and in what order; they do not model how a power failure propagates to clinical records to medication dispensing to patient safety. The rebuilt program did.
Two consumer hardware manufacturers (product reliability)
Attributed field failures in wearable and fitness devices to specific root causes — not aggregate hazard rates. Weibull models over shape and scale parameters distinguished early-life, random, and wear-out failures; the attribution identified which design changes actually reduced failure probability, as opposed to which components happened to fail most often.
Early-stage fintech and health-tech ventures (regulated AI)
Advised on causal AI architecture, interpretability, and ISO/NIST compliance for regulated AI platforms. For platforms where explainability is not optional, designed root-cause frameworks that answered both the operational prediction question and the compliance attribution question from a single model.
The case-study tiers below show the methodology applied across problem types — sometimes drawn from the engagements above, sometimes constructed on synthetic data where client confidentiality is binding. The tier-grouping is honest about methodological depth: how many distinct problems within a domain the framework has been instantiated against.
Tier 1 — Productized engagements
Two domains where the work has reached a deliverable shape: a structural causal layer on top of the existing analytics tooling, a Shiny app the team operates, an Structural Causal Model (SCM) template, and a four-week engagement cycle. The methodology and the surrounding software both ship.
Marketing Mix Modeling (5 pages)
A structural causal layer on top of Robyn or Meridian. The bridge surfaces the back-door confounder regression can’t model and outputs three deliverables: corrected ROIs per channel, per-campaign attribution, and sensitivity-analyzed budget allocation. On a synthetic dataset where truth is known, the bridge corrects TV ROI from $2.83 (Robyn) to $1.84 — within 8% of structural truth ($2.00). On a $50M TV budget, that’s ~$20M of revenue that wouldn’t have materialized.
Enter the section →Supply Chain (3 pages)
SCM models for resilience analysis — how upstream disruptions cascade through tier-1 and tier-2 supplier networks. Strategy and economics framing alongside the methodology. Same engagement shape as Marketing Mix Modeling; productization in progress, with the strategy and methods pages live and the question and components layers still to come.
Enter the section →Tier 2 — Case-study portfolios
Five domains with multiple applied cases each. The methodology has been instantiated against several distinct problems within the domain, demonstrating that it generalizes within the domain even where it isn’t productized into shipping software.
Insurance (5)
Property insurance pricing, utility wildfire risk, attribution, reserving, and the self-insurance decision. The connecting thread is back-door confounding in pricing data — prices are set with information that is also correlated to the loss outcome.
Enter the section →Healthcare (3)
Personalized medicine, statins and hospitalisation, iatrogenic medications. Roadmap items: oncology immunotherapy, treatment-resistant depression, sepsis dynamic treatment regimes, drug repurposing via transportability, adverse-event attribution.
Enter the section →Finance (3)
Bank churn (counterfactual retention), credit risk, and M&A due diligence. The connecting thread is selection bias — customers, borrowers, and acquisition targets are not randomly assigned, and naive regression confuses selection with causation.
Enter the section →Compliance (3)
GDPR, CCPA / CPRA, and NIST CSF 2.0. Each is a Bayesian-network model of compliance risk with pathway decomposition showing where governance controls reduce exposure and which residual risks survive.
Enter the section →Policy (2)
Criminal causation — counterfactual reasoning in legal proceedings, the “but-for” standard made formally tractable. Elective sequencing & mastery — curricular interventions and student outcomes under heterogeneous treatment effects.
Enter the section →Tier 3 — Individual cases
Four domains where a single case study has been built. The methodology applies; the depth is one case. Each is self-contained and uses the same underlying tooling pattern as the other tiers.
Utilities
Utility grid risk — cascade failures and load-shedding decisions under PSPS conditions, with explicit modeling of how upstream conditions propagate to customer-level outage.
Open the case →ESG
Climate ESG risk — physical and transition risks under different IPCC scenarios, decomposed into pathway contributions so a portfolio manager can see which scenario assumptions are doing the load-bearing work.
Open the case →HR
Training effectiveness — counterfactual estimation of skill acquisition. The methodology separates the effect of training from the selection effect of who chose to take it.
Open the case →Regulatory
Regulatory causal evidence — admissibility standards for expert testimony based on causal claims, structured as a decision DAG for evidence weight.
Open the case →Out of scope
The list above excludes a handful of domains where Pearl-ladder methodology either doesn’t fit or where the work hasn’t been done. Pure prediction problems — where no causal interpretation is needed — sit outside the methodology entirely. So do real-time / sub-second decision systems where the inference machinery isn’t fast enough. Image, text, and speech ML problems are typically not where this set of tools earns its keep.
Two adjacencies are worth distinguishing from this list specifically. “Causal AI” used as a marketing relabel of correlational ML is a different thing entirely — the work above always uses an explicit DAG, identifiability checks, and do-calculus reasoning. Difference-in-differences, instrumental variables, and synthetic-control econometrics are methodologically related and structurally similar, but rarely include the explicit DAG and identifiability discipline that Pearl methodology requires.
These exclusions aren’t permanent. They reflect where the methodology has been productively applied so far.
Going deeper
For the practice in one read — positioning, credentials, and the architecture you would be authorizing — see About. For an audience-shaped doorway with three reader profiles, see For Executives. For how engagements begin operationally, see Scope.
If your problem rhymes with anything in the list above, the conversation worth having is which tier the next engagement should aim at — not whether the methodology applies.
info@rung3.ai