Collect claims
Extract employer, title, dates, degree and supplied links. Keep source spans and versions with each assertion.
Proposed GenAI extractionCANDIDATE VERIFICATION COPILOT
Every finding needs evidence.
Inspect the claim. Understand the gap. Choose the next step.
A working interaction, a simulated investigation. Local rules evaluate synthetic source packets. No LLM, ATS or external verification is connected.
Case library
06SYSTEM DESIGN
Bound the tools.
Make the evidence visible.
The prototype runs the policy and review interaction. A production version would add authorized data adapters and carefully bounded AI.
Associate records
Use permitted identifiers to establish which records belong together. A name and school cannot force a person match.
Rules + human clarificationChoose a check
A bounded planner chooses an approved read-only action. Missing access stops a branch; it does not justify a guess.
Proposed agent planningEvaluate evidence
Separate claim differences, actual conflicts and missing sources. Validate references and source independence.
Fixture checks run hereApply policy
Compute review priority and evidence coverage independently. Known conflicts remain visible when another source is missing.
Implemented locallyKeep a human owner
Let reviewers clarify, resolve and escalate with reasons. Log transitions; never turn a priority into automatic rejection.
Local review interactionWhere GenAI earns its place
Unstructured extraction, semantic comparison and cited summaries are useful AI tasks. Exact lookups, access decisions, policy and audit transitions belong in code. Add agent planning when the next check depends on earlier evidence, and compare it with a fixed workflow first.
What a pilot would need
Authorized ATS / credential adapters, an approved model provider, authentication and access controls, a persistent evidence store, reviewer ownership, PII retention rules, adversarial tests and measured outcomes. Suggested stack: FastAPI, Pydantic, LangGraph and a relational database.
Evaluate the whole workflow
Compare manual review, rules-only and rules + GenAI under the same permitted evidence. Measure false escalations, citation support, correct abstention, reviewer disagreement and total time including correction. Fraud recall requires independently confirmed labels. This demo has no fraud-accuracy or time-saving claim.
Technical references
FBI / IC3: layered identity verification ↗
OWASP: prompt injection ↗
PROJECT NOTES
A README you can use.
Scope, walkthrough, policy, implementation and evaluation.