What we have
Findings, live on the platform
A real run: 82,169 claim lines analyzed under one governed tenant, findings ranked and scored against known ground truth.
VForce · gold.fwa_scorecard · tenant: bop
The same governed platform that catalogs and orchestrates your data runs the analysis. Flow ingests the claims, the lakehouse governs the medallion model and Medicare benchmarks, Askura lets you ask the findings in plain language, and a three-layer engine of rules, machine learning, and human feedback finds improper payments, all inside a closed-gap boundary with zero data egress.
A real run: 82,169 claim lines analyzed under one governed tenant, findings ranked and scored against known ground truth.
Every stage is a platform capability, not custom code. The same primitives serve any claims or records client.
Secure file intake (JEFS / SFTP / CSV), schema validation, medallion build.
Canonical claims model, Medicare reference, RBAC, masking, audit.
Rules, unsupervised ML, and benchmarking produce ranked findings.
Analyst validates every AI-assisted conclusion; decisions become labels.
Word and Excel deliverables plus a detailed, traceable audit file.
The platform lands and governs the claims so that every finding is reproducible and every access is policed.
Connectors receive the claims file, validate it against the canonical schema, and build the medallion (bronze to gold) on a schedule or on arrival. External reference data, such as CMS fee schedules by year, is ingested and versioned the same way.
Institutional and professional claim lines (837 / UB-04 / CMS-1500) mapped to one governed model: diagnoses, procedures, modifiers, revenue codes, providers, and charges, ready for analysis.
Billed charges compared to locality-adjusted Medicare rates by service-date year. Relationship, role, and attribute-based access, dynamic masking, small-group suppression, and a tamper-evident audit, on GraphQL, REST, or ODBC.
The conversational layer of the suite. Ask questions of the governed findings in plain language and get cited, human-reviewable answers, with agentic assist for reviewers.
askura.ai →Rules carry the high-confidence findings, machine learning finds the unknowns, and the reviewer makes the system smarter with every decision.
Duplicates, upcoding, unbundling, provider over-use, above-Medicare charges, and impossible timelines, each flag naming the claim, the rule, and the dollars. Fully explainable and reproducible.
Anomaly detection over provider and procedure profiles ranks the borderline cases and surfaces novel schemes the rules were never told to look for, with no labeled data required.
Every reviewer confirm or reject becomes a label that calibrates thresholds now and trains supervised models over time. The engine improves with use, per tenant.
The analysis runs where the data can never leave. Nothing is sent to a third party or used to train anyone's model.
An internal validation ran the engine over a synthetic 82,000-line claims dataset with planted fraud. Recall is measured against the known planted cases.
| Detection | Tier | Recall |
|---|---|---|
| Exact duplicate claims | detected | 100% |
| Extreme Medicare multiples | detected | 100% |
| Phantom same-day volume | detected | 100% |
| Unbundling | detected | 100% |
| Upcoding | detected | 100% |
| Impossible timelines | detected | 100% |
| Moderate over-billing | maybe | 100% |
| Modifier misuse | maybe | 100% |
Synthetic validation with planted ground truth, not client data; near-duplicate detection is on the roadmap for the machine-learning layer. Results on real data depend on data quality and the agreed methodology.
Integrity systems at real scale, brought to government claims analysis.
Trust, safety, and integrity systems for a global online marketplace and its health offering, detecting abuse across high-volume event streams in a closed-loop, event-driven architecture.
Provider master data management and claims-data modernization for a major commercial health payer in a HIPAA-regulated environment.
Senior engineering on claims-organization technology for a national insurance carrier, including claims data structures and leakage signals at high volume.
FWA is one solution on VForce. The same Flow, lakehouse, Askura, and machine-learning primitives serve any claims or records mission.