BehaviourLens AI

A chatbot interviewer reads the transcript and generates a scripted reply, with a noticeable gap between turns. BehaviourLens AI is a voice interviewer that analyses three behavioural dimensions at once — acoustic emotion (pitch, energy, speech rate), visual micro-expression (gaze stability, posture, authenticity) and linguistic sentiment (hedging language, technical markers). All three are fused in a single sub-400ms pass, so it can follow up, encourage or clarify at conversational speed rather than at form speed.

It consolidates three things that are normally three separate steps: evaluating the answer, profiling the candidate's psychometric state, and deciding what to say next. Sequential designs need three or four steps per turn and leave ten seconds of silence in between, which changes how candidates behave — they fill the gap, or they tighten up. Consolidating into one pass gets internal processing under 400ms and keeps the cadence conversational, which is the condition the behavioural reading assumes.

Two layers. The visual forensics module samples webcam frames for active phone use, note reading, tab switching detected through screen reflections in eyeglasses, and additional faces in frame — each behind a high confidence threshold, so a phone on the desk is not a flag and reading from one is. The second layer is the fusion itself: textbook-perfect answers arriving with high nervousness, low gaze stability and a sustained downward reading pattern is a contradiction no single stream would catch, and it marks the session for human review.

No. The platform is built for air-gapped on-premise deployment. Speech-to-text, audio emotion, turn detection, the vector index and the database all ship pre-packaged in the containers, so it runs inside a private data centre, a VPC or a classified network with no outbound dependency.

While the candidate is still speaking, the speculative stream watches the partial transcript and fires an asynchronous vector search on every 25 characters of new text, retrieving the relevant job-description requirements, expected answers and résumé context. By the time voice activity detection confirms the end of the turn, the context for the next question is already loaded — which removes the two-to-three second search a conventional system pays between every turn.

A weighted formula calibrated against human interviewer benchmarks. Technical accuracy carries 60%, evaluated against RAG ground truth. Behavioural consistency carries 20%, measured as the standard deviation of per-question scores, so steady performance is rewarded over erratic swings. Improvement trend carries 10%, comparing the first half of the session to the second to separate candidates who warm up from those who fatigue. Emotion stability carries the last 10%. The output is tiered — Strong Hire, Hire, Consider, Reject — and every component is evidenced in the report.

That is what the five-decision framework is for. NEXT_QUESTION moves on when the answer was good. ASK_FOLLOWUP probes when it was shallow. ENCOURAGE offers support when the forensics show nervousness or hesitation. CLARIFY asks again when the answer was ambiguous. WRAP_UP closes the session. A safeguard caps consecutive non-consuming decisions at two per question, so a candidate cannot be probed indefinitely on one topic.

Webcam frames are analysed in memory and are not written to disk by default — what persists is the structured forensic reading, not the image it came from. Audio is encrypted at rest with per-tenant keys and purged on a retention policy you configure. Each tenant gets its own isolated schema rather than a shared table, and every session carries a cryptographic signature so a scoring decision can be reconstructed and attributed later.