Focus / Teramind: Скрининг

Model it as a rules-aware video pipeline: detect people/actions/objects, attach time/staff/kitchen policy context, evaluate protocol violations with a deterministic rules layer or constrained LLM, and use human review plus monitoring for edge cases.

Start by separating perception from policy. The perception layer extracts structured events from video: people, uniforms, handwashing, knives, hats, restricted zones, timestamps and object interactions. Use lightweight CV models for frequent online checks and stronger VLMs or humans for bootstrapping labels and hard cases.

The policy layer should not be hidden inside a generic prompt. Store kitchen-specific rules, shift schedules and staff permissions in a structured rules engine or retrieval-backed policy store. Convert perception events into facts, then evaluate rules such as “staff member X must wear item Y in zone Z during shift T” or “knife must be washed after raw-food contact.”

LLMs can help map natural-language policy documents to structured rules, summarize incidents and assist review, but the production decision path needs traceability. For deployment, sample frames/clips intelligently, keep latency and privacy constraints explicit, log evidence clips for violations, and measure precision/recall by rule type, camera angle, lighting and site.

First turn “suspicious” into operational categories and severity levels, measure human agreement, build a labeled review set with adjudication, then optimize risk-calibrated precision/recall and downstream business outcomes.

If humans disagree, the first task is not model selection; it is label design. Define categories of suspicious behavior, severity, required action and non-goals. Measure inter-annotator agreement and keep a third-party or senior-review adjudication process for hard examples.

Evaluation should mix ML metrics and product metrics. Offline, use a stratified video set across stores, camera positions, time of day and traffic level. Track precision, recall, false alarms per hour, missed high-severity incidents, calibration and performance by subgroup or environment. If labels remain ambiguous, report soft labels or agreement-weighted metrics rather than pretending there is one perfect ground truth.

Online, measure analyst workload, alert acceptance rate, time to incident review, customer/store outcomes and appeal rate. Thresholds should be risk-based: high-confidence severe events can alert immediately; uncertain low-severity events can go to passive review or sampling. Monitoring must watch drift by store layout, seasonality, camera changes and policy changes.