Keryx AI
Listens to every call. Decides none of them.
Keryx AI adds a layer of machine assistance across the Keryx suite. It screens out nuisance calls before they ring, transcribes and translates live, reads caller stress, fills the incident record as the conversation unfolds, and turns your call history into forecasts. Every output is a suggestion a person can see, question, and override.
What it does
Emergency centres run on attention, and attention is the resource under the most pressure. A meaningful share of inbound traffic is misdials, pocket calls, and automated dialers, each one consuming a call taker for the seconds it takes to confirm there is nothing there. Meanwhile the calls that matter arrive in a language the call taker may not speak, from a person whose state of mind is hard to read, with details scattered through a frightened conversation. And the data that could explain all of it — where calls come from, when, and why — sits in a spreadsheet that gets opened once a quarter.
Keryx AI works across Core, Call, Dispatch, and Message rather than beside them. It scores a call before it rings and steers likely nuisance traffic into a low-priority queue that a person can pull from at any time. On live calls it transcribes and translates in real time, surfaces stress and urgency signals as prompts, extracts addresses, descriptions, and hazards into the record, and drafts the hand-off summary. Across the whole history it forecasts demand by hour and area and flags patterns worth a supervisor's eye. What it does not do is act alone: every suggestion is labelled as one, explained, logged, and one click from being dismissed.
Capabilities
Nuisance and spam call screening
Robocalls, misdials, pocket dials, and known abusive numbers are scored before the first ring and routed to a low-priority queue. Nothing is dropped — a call taker can pull any screened call in one action, and every screening decision is logged.
Live transcription and translation
A running transcript appears beside the call as it happens, in the caller's language and the call taker's. Non-native callers are understood without waiting for an interpreter line, and the transcript becomes part of the record.
Caller state signals
Voice and text are analysed for stress, urgency, and escalation. The call taker sees a prompt — “caller distress rising”, “possible intoxication”, “child on the line” — never a verdict. Signals are advisory and configurable by agency policy.
Records that fill themselves
Addresses, vehicle descriptions, weapons, injuries, and hazards are lifted from the conversation into the incident record as they are said. If a spoken address disagrees with the plotted location, the mismatch is flagged immediately.
Duplicate detection and instant summaries
When several callers report the same event, Keryx AI proposes a merge before the room dispatches twice. At hand-off, a two-line summary is drafted for dispatch and responders, ready to accept or edit.
Forecasting and analytics
Predicted call volume by hour, day, and area for staffing and surge alerts. Hotspots, repeat locations, response-time drivers, and protocol adherence surfaced in dashboards built from the same records the room already keeps.
Standards and interoperability
Built on the open specifications the emergency services community has already agreed on, so it connects to what you run today and what your network is migrating toward.
- NENA i3
- Consumes call, location, and media data from Keryx Core, so analysis works from the full next-generation payload.
- EIDO
- Writes extracted details and summaries back into the incident as structured data other systems can read.
- Human-in-the-loop
- No output changes a call's handling without a person's action. Suggestions are labelled, explained, logged, and reversible.
- NIST AI RMF
- Model development, testing, and monitoring follow the NIST AI Risk Management Framework’s govern–map–measure–manage cycle.
- EU AI Act
- Designed against the obligations for high-risk AI systems used by emergency services, including documentation, oversight, and logging.
- Data residency
- Inference runs in the agency’s own cloud tenancy or on premises. Call audio and transcripts do not leave the agency’s control.
Why centres choose it
— Suggests, never decides
Every screening, flag, and summary is presented as a suggestion with its reason attached. The call taker stays in charge of the call, and the audit log shows what was suggested, what was accepted, and what was overruled.
— Learns your district, stays in your district
Models are tuned on the agency’s own call history and deployed inside its own infrastructure. The centre gets an assistant that knows its geography and vocabulary without sending a single recording to a third party.
— Pays for itself in nuisance calls
Screening alone hands back the minutes a room loses to empty lines every hour. The transcription, extraction, and forecasting come on top of a return the centre can measure in the first month.
Part of the Public Safety suite
Run Keryx AI over a week of your own calls.
Bring anonymised recordings from your centre. We will show you what it would have screened, transcribed, flagged, and forecast — and, just as importantly, where it would have stayed quiet.