A full-stack compliance monitoring platform for Global Systemically Important Banks. Automated control testing, LLM-powered evaluation, and continuous surveillance — built from the model layer up.
Every GSIB runs the same playbook. Massive control inventories. Tons of manual sample analysis and data work. Human reviewers drowning in evidence. And at the end, findings that could have been caught in minutes.
Control testing at major banks runs on spreadsheets, email chains, and copy-paste workflows. A single test of design takes days. Operating effectiveness testing takes weeks. Repeat for hundreds of controls across dozens of domains.
Different reviewers apply different standards. Workpaper quality varies wildly. Senior reviewers spend more time fixing documentation than analyzing controls. Institutional knowledge walks out the door with every departure.
Controls are tested on a cycle — quarterly, annually. Between tests, exceptions go undetected. By the time findings surface, the damage is done. Continuous monitoring exists in theory but rarely in practice.
Real screens from the production platform. No mockups. No Figma. Working software.
AI-generated issue identification with root cause quality scoring, remediation tracking, validation evidence assessment, and theme analysis.
10 surveillance domains with health score gauges, rule counts, exception tallies, and pass rates. Wire Transfers through Trade Surveillance.
AI-powered evaluation with composite scoring gauge, seven testability dimension bars, and structured analysis with strengths and gaps.
Full exception population browser with severity-coded rows, OFAC matches, structuring patterns, and review status tracking.
Four integrated modules — Monitoring for continuous surveillance, Control Testing for design analysis, operating effectiveness, test of one, and PRC inventory management, Issues for AI-driven root cause and remediation, and Reports for regulatory-ready output.
Full-population and sample-based monitoring with AI-powered exception triage. Populations ingest automatically, rules fire against live data, and exceptions route to reviewers with AI-generated severity scores.
Automated control evaluation across the full testing lifecycle. The LLM evaluation engine scores controls across seven testability dimensions, generates Test of Design analyses, and produces structured workpapers.
AI-generated issue identification from test results. Root cause analysis quality scoring, remediation plan tracking, validation evidence assessment — with full lifecycle management from discovery to closure.
Control health dashboards, issue aging analysis, and automated report generation. Export to PDF, Excel, and Word. Regulatory-ready output formatted for executive and committee presentations.
PathLighter doesn't use AI as a wrapper. The LLM evaluation engine is built on a 950+ normative control inventory organized across 10 novel archetype classifications — each with domain-calibrated rubrics that took hundreds of hours to tune. Calibrating an LLM to evaluate compliance language with precision is one of the hardest problems in AI governance, and it's the core of everything here.
Controls are evaluated across seven testability dimensions — specificity, measurability, frequency, evidence quality, ownership clarity, exception handling, and automation level. Scores aggregate into a composite testability rating with weighted importance.
The hardest part of the entire system. The evaluation engine is calibrated against financial compliance language, regulatory frameworks, and 10 distinct control archetype patterns — each requiring different rubrics, scoring weights, and evidence expectations. Getting an LLM to reliably distinguish an APPR control from a RECON control, and evaluate each against the right standard, required extensive iteration and domain expertise that can't be shortcut.
Systematic bias detection across LLM outputs — position bias (first-option preference), verbosity inflation (longer = better scoring), and self-enhancement (AI favoring its own prior outputs). Each bias vector is measured and mitigated.
Structured workpapers generated automatically from test results. Formatted with evidence citations, finding classifications, and recommendation language — all consistent with PCAOB and regulatory standards.
Interactive walkthrough assistant that gathers context about control environments through structured conversation. Identifies gaps in control descriptions, asks targeted questions to fill them, and tracks which testability dimensions have been addressed.
AI-powered severity scoring for monitoring exceptions. Each exception is evaluated against historical patterns, control context, and risk impact — then routed to the appropriate reviewer with a pre-generated analysis brief.
The architecture under PathLighter — purpose-built systems for compliance automation, continuous monitoring, and LLM governance at institutional scale. We've loaded 10 million+ transactions to demonstrate real-time performance at production volumes.
Full-population surveillance engines with rule-based and AI-powered anomaly detection — exception queues, health score gauges, threshold alerting, and sample-based reviews at scale.
Structured LLM-as-a-Judge scoring that replaces subjective review with quantifiable assessment — seven testability dimensions, weighted aggregation, and human-in-the-loop governance.
End-to-end automated testing from control description analysis through workpaper generation — Test of Design, Operating Effectiveness, and walkthrough analysis with AI-assisted evidence assessment.
A normative inventory of 950+ controls organized across 10 regulatory archetypes — each with domain-specific evaluation criteria, population definitions, and expected evidence standards.
AI-generated issue identification triggered from test results — with root cause quality scoring, remediation tracking, validation evidence assessment, and theme analysis across populations.
Automated generation of committee-ready reports — control health scorecards, issue aging analytics, remediation velocity metrics, and export to PDF, Excel, and Word.