TruStory
A proof of concept for the Mississippi Department of Public Safety that cross-checks recorded interrogation statements for contradictions, with every flag traceable to its source statement, speaker, and timestamp.
TruStory is a proof of concept built through the Mississippi AI Innovation Hub (Fall 2026 cohort), for the Mississippi Department of Public Safety and the Mississippi Bureau of Investigation. Team WeThree’s proposal was accepted into the cohort on September 22, 2026. Work on the proof of concept has not started yet; this page will be updated as it progresses.
The problem
Investigators review recorded interrogation statements by hand, with no systematic way to compare what one subject says against other statements in the same case file.
The plan
Interrogation video is transcribed with timestamps and speaker diarization, statements from a case are grouped together, and a comparison engine flags likely contradictions using LLM-assisted claim extraction, natural-language-inference classification, and sentence-embedding similarity. Every flagged item is meant to cite its source statement, speaker, and timestamp. The tool is designed to be advisory only, with a human always in the loop, an append-only audit log, and no credibility scoring or behavioral analysis.
My role is AI and evaluation lead: selecting the speech-to-text and diarization models, building the claim-extraction and contradiction-detection logic, and designing the evaluation framework. The proof of concept is planned to run on 10 synthetic test cases before any real presentation at the Hub’s showcase.
No results exist yet. Numbers will be added here once the proof of concept actually produces them.