VerdictX

Inspiration

The numbers stopped us cold. Racial minority applicants are denied mortgages 1.7 times more than white applicants. Racial minority patients have insurance claims denied 50% more often. Algorithmic systems make these decisions in milliseconds with zero human review, and most people never appeal not because they don't have a case, but because they don't have a lawyer.

We asked: What if the adversarial structure of a legal appeal could be fully automated?

What We Built

VerdictX is a multi-agent adversarial AI system that fights algorithmic denials — loans, insurance claims, housing applications, and benefits terminations. Paste your denial letter. Five specialized agents simultaneously attack it from every angle. A Denial Defender must justify the algorithm's decision using real regulatory citations. When it concedes two or more arguments it cannot rebut, the override fires and produces a ready-to-send formal appeal letter grounded in actual case law and federal statute. The mechanic mirrors appellate jurisprudence: the defense doesn't lose because the attack wins, it loses because the defense collapses.

How We Built It

All agents are orchestrated through IBM Orchestrate across four sequential waves. The parser runs first, then four attacking agents run in parallel, then the Denial Defender rebuts all four, then the Override Judge decides, then the Appeal Letter Writer generates the output.

Each agent has a locked persona, a strict JSON output schema, and a domain-specific knowledge base sourced from real regulatory documents.

The Bias Auditor's knowledge base contains the CFPB's own BISG proxy methodology for detecting race and national origin discrimination via surname and geography, the Interagency Fair Lending Examination Procedures, and CFPB Fair Lending enforcement reports. It hunts proxy discrimination signals — surname, zip code, income type, and algorithmic decisioning flags.

The Precedent Agent's knowledge base contains major CFPB and DOJ enforcement actions, including CFPB v. Ally Financial, United States v. Trident Mortgage, and CFPB v. Citibank, plus the Anderson v. United Finance circuit court decision establishing adverse action notice standards. It finds comparable cases where identical denials were overturned.

The Circumstance Agent's knowledge base contains the CFPB Credit Invisibles Report documenting that 26 million Americans are invisible to scoring models, the CFPB Alternative Data RFI recognizing rent and utility payments as valid creditworthiness evidence, and Federal Reserve research on FICO's structural blind spots. It argues everything the algorithm structurally cannot see — life events, alternative credit evidence, thin file status.

The Legal Agent's knowledge base contains the actual statutory text of ECOA 15 U.S.C. 1691, FCRA 15 U.S.C. 1681m, Regulation B 12 CFR 1002.9, and CFPB Circulars 2022-03 and 2023-03 establishing that algorithmic complexity is not a defense against adverse action notice requirements. It identifies specific statutory violations with section-level citations.

The Denial Defender's knowledge base contains Fannie Mae Selling Guide Chapter B3, the OCC Comptroller's Handbook on underwriting standards, and the Supreme Court's Inclusive Communities decision establishing the robust causality requirement for disparate impact claims. It must justify the denial using these sources or concede.

The Override Judge has no knowledge base; it is a pure logic layer that audits every rebuttal for quality. A rebuttal without a specific citation is automatically classified as a concession. At two or more concessions, the override fires. The Appeal Letter Writer then receives only the conceded arguments — arguments the Defender successfully rebutted are explicitly excluded, ensuring the letter is airtight.

Stack

IBM Orchestrate handles all agent orchestration and wave sequencing. We parse photographed or handwritten denial letters into structured JSON. OpenAI powers supplementary agent reasoning. ElevenLabs provides a voice interface for submitting denials and receiving appeal summaries. Twilio handles automated outbound calls connecting users with consumer rights lawyers. Auth0 manages OAuth 2.0 authentication. MongoDB stores denial data, debate transcripts, and appeal letters. DigitalOcean handles hosting and deployment.

Challenges

Every agent needed verbatim regulatory text, not paraphrased summaries; agents are only as credible as their citations. Early Denial Defenders produced rebuttals that sounded valid but didn't address the specific argument raised. Requiring every rebuttal to cite a specific document and section solved this cleanly — a rebuttal without a citation is automatically a concession. Calibrating the override threshold to two or more concessions mirrors the legal standard for a prima facie case, multiple independent grounds of failure, not a single arguable point.

What We Learned

Heterogeneous agents with locked antagonistic personas dramatically outperform a single agent asked to consider all angles — backed by Du et al. at ICML 2024 and the A-HMAD paper in Springer 2025. The problem is also larger than we expected. 765 million dollars per year in excess mortgage interest is charged to minority borrowers. Black applicants are half as likely to receive unemployment benefits per GAO findings. Algorithmic accountability is one of the largest unaddressed civil rights problems of the decade.

AI Usage Disclosure

Claude (Anthropic) was used for ideation, agent prompt engineering, knowledge base sourcing, and architecture planning. OpenAI powers specific agent reasoning tasks. IBM Orchestrate manages all orchestration. Application logic, UI, database schema, integrations, and deployment were built by both the team and Claude Code.