Deepfake detection built for auditable decisions
Jarkata normalizes conference and browser media into one detection engine, records every outcome in one schema, and produces alerts and evidence without hiding connectivity or model failures.
No published accuracy or latency claim is made until measured on a labelled evaluation corpus and target devices.
Implemented architecture
Six capture paths. One contract. One record.
Zoom RTMS
Tier 1 adapter
Microsoft Teams
Tier 1 adapter
Google Meet
Tier 1 adapter
Webex
Tier 1 adapter
Browser extension
WebRTC/JPEG path
File replay
Video test path
Operational truth, not a mock result
Clean is supplied by the store and excludes unavailable records. It is never calculated as total minus flagged.
Verified in the repository
What the current build actually does
Normalized capture contract
Frames, PCM audio, and source metadata stay separate from detection logic.
Xception ONNX inference
The checked-in model is loaded once and receives 299×299 preprocessed frames.
Honest failure states
Unavailable analysis is never relabelled as clean; reduced connectivity remains visible.
Evidence controls
Zero-recording mode defaults on; reports can be generated without retaining raw media.
Operational outputs
OS notification, Slack/webhook, SIEM, CSV, HTML reports, and evidence packages are implemented.
Drift and route monitoring
Cloud/on-device usage, measured RTT, timeouts, and prediction drift have dedicated metrics.
Work that remains intentionally labelled
Evaluate against your threat model
Pricing and service-level guarantees are not published until deployment scope, evaluation data, and operating requirements are agreed.
Plan your capture path