AI-Driven Psychological Profiler for Criminal Profiling
Criminal profiling has, for most of its history, been a craft. A handful of seasoned investigators read patterns in case files the way doctors used to read symptoms — instinctively, with experience as the only real teacher. The work was effective, but it was slow, expensive, and rarely scaled to the volume of data modern law-enforcement agencies actually deal with.
Over the past year, our team at EV Lab has been quietly building something we think changes that equation: an AI system that ingests case data, statements, digital footprints and behavioural records, and produces a structured psychological profile that an investigator can use as a starting point — not a verdict. This post is about what we built, what we learned, and where we think this technology is heading.
Why we built it
The team got pulled into this problem after a series of conversations with state law-enforcement units. The pattern was always the same: a small specialist cell, often one or two officers, drowning in unstructured data — call records, chat logs, financial trails, forum posts — with no way to surface the behavioural signals hidden inside them. Profilers existed, but they were rare, expensive, and rarely available outside major cases.
"We don't need an AI that solves the case. We need an AI that lets us read 10,000 pages in two hours instead of three weeks."
That conversation reshaped our scope. Instead of trying to build something that produced answers, we built something that produced structured questions — categorised hypotheses about a subject's likely traits, motivations and risk levels, with the underlying evidence cited inline so the investigator could push back on any one of them.
What's actually inside
The system is a pipeline, not a single model. At a high level it does four things:
- Ingestion & redaction. Raw evidence is normalised into a structured graph. Personally identifiable information is automatically tagged so analysts can choose what's used in modelling and what isn't.
- Behavioural signal extraction. A combination of language models and a custom-trained classifier identifies psychologically meaningful events — escalation patterns, fixations, deception markers, social withdrawal, and so on — across the entire corpus.
- Trait inference. The signals roll up into a structured trait profile, scored against established frameworks (Five Factor, HEXACO, dark-triad indicators) with confidence intervals on every dimension.
- Reporting. The output is a navigable report. Every claim is linked back to the specific evidence that produced it, so an investigator can collapse the report and walk it source-by-source.
What worked — and what didn't
The thing that worked best, surprisingly, was being conservative about predictions. Early prototypes tried to output strong, specific behavioural forecasts ("subject is likely to escalate within 30 days"). Field tests showed those were the predictions investigators trusted least, even when they were technically accurate. The version that landed instead surfaces ranked hypotheses with explicit uncertainty, and it gets used.
What didn't work: trying to be domain-agnostic. We initially built the trait inference layer to work across any kind of subject — corporate, criminal, social. The signals are too different. We narrowed scope to specific case categories (cyber, financial, organised crime) and accuracy jumped dramatically.
Things we'd do differently
- Build the audit trail before the model. Investigators want to know why the model said something, not just what it said. We bolted explanations on later — should have been there from day one.
- Design for one-officer workflows. The system was originally built for teams of analysts; in practice most users are solo, and the UI now reflects that.
- Don't underestimate the data plumbing. Roughly 60% of the engineering effort went into ingestion and normalisation, not into the models themselves.
Where this is heading
The next milestone is multi-modal evidence — image, video and audio fed directly into the same pipeline. We're also exploring a domain-specific fine-tune that runs entirely on-premises, since several of the agencies we work with cannot, by policy, send any data to a cloud-hosted model.
If you're working on something similar, or if you have a case-class we haven't seen yet, we'd love to hear from you. Drop us a line at [email protected] — the system gets meaningfully better with every new domain we learn from.
— The EV Lab Research Team
