CertVault: LLM-Assisted Certificate Verification with Rule-Based Decisions and Human Review
Certificates issued on behalf of students from online learning platforms are routinely collected by colleges as evidence of learning, however, validation is done manually and it is relatively straightforward to generate fake certificates in an edited PDF format. In this work, we introduce CertVault, a certificate management and validation system that utilizes a large language model (LLM) as an evidence extraction tool and not as a judgment maker. Uploaded PDF documents are first scanned for malware, structured fields and anomalies are then extracted using Google Gemini under a fixed response schema, and a deterministic rule engine determines the status of the certificate. Automatic validation of a PDF also requires a validation of the same certificate from the issuing organization's publicly available verification site; every case where automatic validation is not possible is sent to human reviewers based on a set of predefined reason codes. From a technical perspective, CertVault integrates multimodal LLM document understanding with schema-constrained structured output, a confidence-thresholded conjunction of decisions, Unicode-token-subset name matching and a prompt injection detection mechanism.
Authors
- Manisha Dhage
- Krish Kumar
- Sanyakta Bhatt
- Neha Kumari
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24351337
- Primary Topic
- Handwritten Text Recognition Techniques
- Type
- preprint