AI call QA for AWS Connect Voice Streams.
Native adapter for AWS Connect's PCM 16-bit 8kHz voice streams. Live transcription, scoring against your own scorecard, mid-call breach detection within 30 seconds, and HMAC-signed webhooks back to your CRM, agent desktop or supervisor screen.
AWS Connect's voice stream in. Compliance evidence out.
AWS Connect publishes voice streams to Kinesis Video rather than to a WebSocket, so unlike our Twilio integration — where we read Media Streams frames directly — this one runs through a small bridge in between. You deploy a single Lambda from our template; it re-emits those frames over our generic WebSocket protocol as PCM 16-bit 8kHz mono, which our AWS Connect adapter consumes natively. It is a well-understood pattern for AWS Connect, and it is the only extra moving part.
Deploy the bridge Lambda
Single-purpose Lambda from a CloudFormation template. Reads from your Kinesis Video stream, re-emits to our WebSocket endpoint with your CallGuard API key. Runs in your own AWS account so audio never leaves until it hits CallGuard.
We score in real time
The adapter parses PCM 16-bit 8kHz frames, runs live transcription with speaker separation, and feeds the rolling transcript into Claude every 30 seconds for breach detection at over 75% confidence.
Webhooks fire back
HMAC-SHA256 signed POST to your endpoint within 30 seconds of a breach being detected, plus a final scored event when the call ends. Compatible with AWS EventBridge so you can fan-out to Connect tasks, agent UI overlays, or a Slack channel.
Built for the way AWS Connect customers actually deploy.
Native PCM frame parsing
The AWS Connect adapter handles raw PCM 16-bit 8kHz frames. No re-encoding step. The bridge Lambda is small, single-purpose and stays in your AWS account.
EventBridge friendly
Webhook payloads are JSON with consistent event names (session.breach_detected, session.scored) so they slot directly into EventBridge rules. From there, fan out to Connect Tasks, Lambda, Step Functions, or third-party endpoints.
30-second breach cadence
Mid-call breach detection runs every 30 seconds against the rolling transcript. Only emits a breach when the AI is over 75% confident a criterion has clearly already failed. No alert fatigue.
Agent desktop overlays
Subscribe to live breach webhooks from your agent desktop (Connect CCP or your custom UI). Show whisper-style nudges to the agent the moment the AI flags a breach.
Per-tenant AI learning
Your supervisors correct AI scores in the dashboard. Corrections feed back into future scoring so the AI calibrates to your interpretation of "fair value", "vulnerable", and "good outcome", not a generic baseline.
Encryption and residency
Audio (if retained) is encrypted at rest with AES-256-GCM. Hosted in the UK (London region). UK GDPR Article 28 DPA published.
Common questions, answered.
Why a Lambda bridge instead of a direct Kinesis Video integration?
Two reasons. First, the bridge keeps audio inside your AWS account until CallGuard explicitly receives it, which is the data-residency story most security teams want to hear. Second, the bridge is small, single-purpose and cheap to run, so you do not need to grant CallGuard cross-account IAM permissions to your Kinesis stream.
How do calls map back to Connect's contact records?
The bridge passes Connect's ContactId as the external ID on the CallGuard call record. Whatever attributes you set on the contact (queue, agent ARN, customer ANI) are forwarded as session metadata and returned in webhooks.
Can we use this with Connect post-call recordings instead of streaming?
Yes. Push the Connect recording URL or audio file into our REST ingest endpoint after the call ends. You get the same scoring output without the mid-call breach webhooks. You can run streaming for compliance-critical queues and post-call ingest for the rest.
Does the bridge Lambda count against our AWS service limits?
The bridge runs as concurrent executions, one per active streamed call. AWS default Lambda concurrency is plenty for most contact centres. For large floors we can switch the bridge to ECS Fargate for tighter capacity control.
What happens if the streaming connection drops mid-call?
The session enters a graceful-disconnect state. Whatever has been transcribed and scored up to that point is preserved. When the bridge re-connects (Lambda retries automatically), the session resumes against the same external ID.
Can we keep audio in our own S3 instead of CallGuard storage?
Yes. The default deployment retains audio in CallGuard for the audit trail, but you can keep the audio in your own bucket and pass a presigned URL to our REST endpoint instead. CallGuard fetches it to transcribe and score, and you keep the master copy.