Skip to main content

Working hackathon MVP · Investigation support

Give investigators a prioritized path through public Medicare data.

Indicium, formerly Integrity, is a working hackathon system that helps investigators prioritize providers for review. Its risk signals trace back to source data; they are not fraud findings.

1st place · ACT-IAC 2026 CMS track

What we’ve demonstrated: Indicium is a company-led ACT-IAC hackathon winner built on public data. The metrics below describe that build and its analytical validation. They do not establish federal production past performance, agency fraud findings, or realized savings.

Inside the current build

Historical screens from Indicium’s hackathon build, then called Integrity, running on its built-in demonstration data. The provider shown is fictional.

Indicium provider profile for a fictional pain-management provider, showing a 94.2 risk score, a bar chart of the top risk signals, a peer-comparison radar chart, the signal mix and an evidence matrix.
Provider risk profile. The score comes with the signals behind it, a peer comparison and the evidence for each factor.
A billing-trajectory chart comparing one provider with specialty peers and the national average, above explained risk-factor cards and a triage snapshot with flags.
Explained risk factors. Every factor states the measurement in plain language, so an investigator can check it.
Fairness Review page showing an overall assessment within acceptable bounds, a 3.8x maximum flag-rate disparity across 14 groups, and flag rate, false-positive rate and average score by specialty.
Fairness review. Flag and false-positive rates are compared across specialties and regions.
Model Validation table showing precision at K against confirmed fraud cases, with 42 percent precision in the top 100.
Model validation. Precision at K against confirmed exclusion and DOJ/OIG cases.
548M
Rows Ingested
1.47M
Providers Scored
42%
P@100
$500M+
Public enforcement-case value

The hackathon build packages ingestion, scoring, and an interactive investigation dashboard in one Go binary. It combined 22 CMS datasets covering 10 years, scored 1.47 million providers, and built an analytical graph with 82.6 million edges from processed public records. Twenty-seven features across eight categories feed composite, machine-learning corroboration, and gradient-boosted-tree scoring methods.

A separate out-of-sample check matched 37 high-ranked non-LEIE providers to public enforcement cases involving over $500M in alleged or resolved fraud. These historical matches are not new agency fraud findings. The reported 6.9-year average lead before OIG exclusion is retrospective, not demonstrated prevention or realized savings. Read the full case study for methodology, backtest results, and out-of-sample validation details.

What you can inspect in the hackathon build

  • Public data only: No protected health information (PHI), data use agreements, or claims-level access required
  • Live weight adjustment: Investigators tune feature weights in real time and watch scores re-rank immediately
  • Full transparency: Every score traces to source data, SQL queries, peer groups, and weights so investigators can inspect how it was produced
  • Single binary: DuckDB, in-process GraphWizard, and all assets compile into one deployable system
  • GraphWizard integration: 6 graph algorithms calculate features associated with shared relationships, prescribing isolation, and exclusion proximity

What Production Still Has to Prove

Before production work, we would agree on authorized public data, the investigative workflow, validation measures, explanations, access controls, accessibility, deployment, and operating responsibilities. The hackathon gives us a starting point. We would still need to test usefulness and controls with the people using and overseeing the system.

What we’re working on next

Pending work as of November 2026. Plans can change as we learn.

  • Planned

    Investigator case management

    Sign-in, a case list, case detail and reports, so reviewers can record what they checked and why.

  • Planned

    Repeatable rankings

    Break score ties consistently, so the same data always produces the same ranked list.

  • Planned

    Security and dependency review

    Review sign-in, data export and third-party packages before any pilot.

  • Planned

    More public context data

    Add neighborhood-level public data alongside the current CMS datasets.