For jail and prison intelligence units
Read every call. Review the ones that matter.
Keyword alerts miss the conversation and bury reviewers in false hits. Call Review transcribes each recorded call, understands it as a whole, and queues the few that need a person, with the exact lines and timestamps.
In development. Call Review is early. What follows is the design.
Pipeline
Privilege first, then analysis, then a person
Privileged calls never reach the model
Numbers registered to attorneys, clergy and other protected parties are blocked before transcription. Matches to the do-not-record list are logged and dropped, and the list is the facility's to maintain.
Whole conversations, in any language, with coded speech in context
The system reads the full exchange rather than scanning for words, so "the thing" and "my cousin" mean what the rest of the call says they mean. Each flag carries the lines that support it, their timestamps and a confidence level.
A queue for people, with the right unit on each item
Threats and witness tampering go to investigations, contraband to facility security, welfare concerns to medical and classification. A reviewer listens to the flagged segment, confirms or dismisses it, and only confirmed findings go anywhere else.
Worked example · synthetic data
One call, from recording to confirmed finding
How one call moves from recording to a confirmed finding. Fictional call, no real people or recordings.
Call 0931-4471 · 11 min 42 s · outgoing to a listed personal number
The facility's phone system recorded the call, as it records every non-privileged call. Call Review checked the number first: not on the attorney or clergy list, so the call was cleared for transcription. The system then read the whole conversation, not a word list, and flagged one 36-second segment. No word on any keyword list appears in it.
What the system noticed
- The caller tells a third party to contact someone he describes by her job, not her name.
- "The fourteenth" is the caller's own preliminary hearing date on the court docket.
- "Works mornings at the store" matches the complaining witness in that case.
- The instruction is to keep her from showing up, with a consequence implied.
A keyword system scanning for "threat," "kill," "hurt" or "witness" would have passed this call. Reading the whole conversation is the difference.
People took it from there
Nothing left the review queue until a person listened. The system's job was to find the 36 seconds worth listening to.
Why it is different
Built for the reviewer, not the alert count
Whole-call reading
Flags are based on what a conversation means, not on a word list, which cuts false hits and catches arrangements that never use a flagged word.
Exact lines and timestamps
Every flag jumps the reviewer to the segment that produced it. Nothing is summarized without the source underneath.
Findings flow both ways
When a reviewer confirms a finding, it goes into the case file in Lead Engine with the recording attached. And when Lead Engine ranks a case, it can ask Call Review whether any of the top leads has come up in a call.
Privilege and audit by design
Blocked calls are never transcribed; every playback and flag decision is logged and assumed discoverable.
Works with existing phone vendors
Designed to sit on the recordings and call records the facility's phone provider already produces. No new equipment.
Measured false-positive rate
The review queue reports how often flags are confirmed, by category, so thresholds can be tuned by the facility.