Seen. Heard.
Connected.

Sifft analyses audio and video together — connecting people, words and events across every source, with every finding traceable to its source. Briefed in plain language. On watch across archive and live. Built in the UK.

Scroll
sifft> did Jessica hear the gunshot? sifft> alert me when a tank appears on camera 3 sifft> everything the commander said about the bridge sifft> who came and went overnight — anything unusual? sifft> find everyone from this bodycam in the warehouse CCTV sifft> map the contact network around this suspect sifft> find scenes where someone hands over a package sifft> notify me if this person is seen on any feed sifft> what was being discussed when the weapon was drawn? sifft> what was said when the alarm went off?

Hours of footage. One analyst.
One timeline scrubber.

Two hundred hours of CCTV from one incident. The options: scrub it by hand, or commission a brittle pipeline that answers one fixed question. When the next question arrives — was there a red car? did they meet before? — you start over.

So footage gets searched once, badly, then never again. The audio is rarely touched at all. And live feeds are watched by motion sensors that cry wolf, or by people who can't watch forever.

200 hrs of CCTV from a single incident is routine
Weeks to review it manually, question by question
Once how often most footage is ever searched

Beyond search.
Into investigation.

Search tools retrieve clips. Sifft is an audio-visual evidence exploitation and intelligence platform — it connects what was said to who appears, where, and when, across every source in a case. Investigation, not retrieval.

Ingest  →  Analyse  →  Correlate  →  Monitor lines of inquiry  →  Traceable intelligence
01
Perceive

Every frame, every voice, every plate

Uploads or live streams. Open-vocabulary detection finds whatever you describe in plain words — no retraining, no fixed classes — alongside tracking, re-identification and word-timestamped transcription. Gunshots, alarms, raised voices: always heard, whether or not anyone asked.

Video · Audio · Imagery · Live & archive
02
Evidence

A permanent, auditable memory

Everything observed becomes structured evidence — entities, timelines, and a knowledge graph of relationships: who met whom, where, what was said, what came next. Full provenance on every record. Same footage, same evidence, every time. Nothing is invented.

Entities · Timelines · Knowledge graph · Provenance
03
Investigate

Agents that answer, and stand watch

Ask in plain language — "who did she meet between 3 and 5?" — and get cited, verifiable answers. Task agents like analysts: profile a subject, map a contact network, deliver the morning brief. On live feeds, standing watches raise verified alerts, not noise.

Natural language in · Cited answers out
Archive and live are the same system. A question you can ask of last year's footage is a watch you can set on tonight's.
Cross-modal by design. What was said, linked to who appears, where — across CCTV, bodycam, aerial and audio. One picture, everything seen and heard.
Case file · Demo The knowledge graph. Relationships between people, objects, events and places — held as evidence, not inference.
co_appears · 147 interacts_with interacts_with precedes · +12s near speaks · heard→seen precedes · +15m Jessica Subject_A Handover · 14:32 White van Gate 4 “…meet at the loading bay” Gunshot · 14:47
Hover or tap any node or relationship — every edge carries its provenance.
● seen ○ heard ◆ event / place — relationship ┅ cross-modal

Built for the sharp end.
Wherever footage outruns people.

M·01
Lead mission

Live site protection

Standing watches over live site feeds — "alert me if anyone approaches the east fence after dark" — verified before they reach an operator, evidence attached. A briefed guard force, not a motion sensor.

M·02

Standing briefs

"What happened overnight? Who came and went? Anything unusual?" — surveyed, compiled and delivered to the morning shift, unprompted, every day.

M·03

Post-incident investigation

Ingest the archive once; answer every question in seconds. Follow a person across cameras, link what they said to who they met, build the dossier — cited back to frame and timecode.

M·04

Force protection

Persistent understanding of activity around bases, convoys and installations — patterns of life, repeated visitors, anomalies.

M·05

Policing & CT casework

Cross-camera, cross-case search over bodycam, CCTV, seized media and intercepted audio — who, where, when, with whom, saying what — at evidential standard.

M·06

Critical infrastructure

Continuous monitoring of high-consequence sites where a missed event is not an option and cloud connectivity is not a given.

Agentic where it helps.
Accountable everywhere.

T·01

Provenance on every finding

Originals preserved untouched. Every finding links to source and timecode; observed fact stays distinct from machine inference; every analysis records how it was made, so an investigator can verify any result — byte for byte, months later.

"Can you reproduce this?" — Yes.
T·02

Humans decide what matters

Standing investigations run to an authorised brief — defined objectives, controlled sources, escalation rules, audit trail. Never open-ended surveillance. Watchlists and alert rules are authored by operators, never by the system.

The AI helps you find things. It does not decide what is worth finding.
T·03

Deploys where you operate

Cloud, on-premise or fully air-gapped — one codebase. Evidence never leaves your estate: no external cloud, no third-party APIs, no licence that expires mid-operation. Information control, by architecture.

Sovereign by architecture, not by promise.
UK-founded Dual-use by design In dialogue with UK defence & CNI operators Open-source model stack Working prototype · live capability in development
Contact

See it run on
your footage.

We brief defence, security and critical-infrastructure teams — and the investors who back them. Bring representative footage and we'll show you what Sifft finds.