Enterprise GPU server
Dell PowerEdge or HP ProLiant, NVIDIA GPUs
The standard deployment. One rack server runs both tiers for a site; add nodes as cameras grow.
- Dual Xeon, ECC memory, NVMe
- NVIDIA data-centre GPUs
- Rack or tower
Agentic video intelligence, on your premises
Your cameras see everything.
EnlytiX tells you what matters.
EnlytiX reads every frame from the cameras you already own, describes what is happening in plain language, decides whether it matters, and acts: an alert to the right person, an answer to a question, a report on the desk by morning. No footage leaves your site.
- 0new cameras to buy
- Secondsto find an event, not hours of scrubbing
- Plain languageto add a new incident type
SenseWatching 9 cameras
ThinkWaiting for something worth reading
ActNo action yet
The problem it solves
99%
of recorded footage is never reviewed. Incidents are found after the fact, if at all, by someone scrubbing a timeline.
1 rule
per thing you want to catch, in classic video analytics. Every new hazard means a new model, new training data and a new project.
Cloud
video AI means streaming your floor, your workers and your faces off site. Most factories and regulators will not accept that.
Two tiers, one system
EnlytiX sits on top of a perception layer that already runs on your cameras. Tier 1 sees; Tier 2 understands.
Tier 1
Human Analytics
Fast perception running next to your cameras and recorders.
- People and vehicle detection with tracking
- Face recognition against your enrolled list, unknown-visitor clustering
- Line crossing in and out, occupancy and crowding
- Number-plate reading at gates
- PPE checks, unattended objects, thermal readings
- Sensor events from recorders: motion, audio, tamper
Tier 2
EnlytiX
A reasoning layer that reads each event the way an experienced supervisor would.
- Zero-shot incident classification, defined in plain language
- Written incident analysis: why it was flagged, what to do
- An agent you can ask questions, in English or বাংলা
- Search all footage by description or by photo
- An incident graph that finds patterns across cameras
- Reports, notifications and a full audit trail
Sense, think, act
Sense
Every camera and recorder feeds in: video, snapshots, and sensor events such as motion, audio exceptions and tampering. Existing cameras, existing recorders.
Think
Each event is read in context: which camera, which area, what that area is watched for, what time it is. The system writes one honest sentence about the scene and decides if it is Normal, Monitor, Alert or Critical.
Act
Alerts go to the right people on the dashboard, mobile app and email, or to on-site sirens and strobes. Everything is indexed so it can be searched, asked about and reported on.
Everything in the walkthrough
Each item opens the section where you can try it.
Live operations
This is what an operator sees. Events arrive from the cameras, are read within seconds, and are filed by severity. Only what clears your threshold interrupts anyone.
All events, all sources
Every event is counted, classified or not, so these totals match each source’s own view. Line crossings, crowding, anomalies and recorder events sit on one timeline.
Last 7 days by source
Busiest cameras
Classified events by severity
Indexing status
How far the reading and search index has caught up with the cameras.
Describe it, and it is watched
Classic video analytics needs a trained model for every hazard. EnlytiX reads each frame as a scene and checks it against incidents written in plain language. Adding one is typing a sentence. It can also answer “nothing is wrong”, which is what keeps false alarms down.
Add a new incident type
Write what you want watched for, the way you would brief a guard. Pick the kind of camera it applies to and how serious it is.
No retraining. No new model. No downtime.
Write an incident on the left and activate it. You will see how it is added to that camera type and which frames in the sample library it would have caught.
Each kind of camera is watched for its own things
A gate camera should not be asked whether an aisle is blocked. Cameras are grouped by type, and each type has its own incident list and its own yes/no checks. Universal hazards such as fire and a person down are watched everywhere.
Ask your site a question
The agent answers from your own incident records, counts and footage. It shows its working, cites the frames it used, and says so when it cannot check something. Type, or speak in English, বাংলা, or a mix.
Hello. I can see every camera on site and every incident on record. Pick a question on the right or ask your own.
Don’t watch video. Search it.
Every event is indexed by what it looks like and what it means. Describe a scene or a person, or start from a photo, and get ranked matches across every camera in seconds, however large the archive grows.
People and identity
Know who is on site, spot who should not be, and follow a person across cameras, with privacy controls that decide who sees faces at all.
Face recognition against your enrolled list
Each capture is matched against enrolled photos. Matches arrive as side-by-side cards; low-confidence matches are marked for review instead of being trusted.
Unknown visitors, grouped automatically
Faces that match no one are clustered, so one stranger seen nine times is one visitor, not nine alerts.
Following one person across cameras
Select a camera to see when the same person passed it.
Find people by what they wear
No face needed. Describe clothing and colours.
Enrolling someone
- Add a photo from HR records or a capture already on file.
- Assign a list: staff, contractor, visitor, watchlist.
- Done. The next pass in front of any camera is matched. No restart.
Privacy by design
- Everything runs and is stored on your premises
- Search and reasoning work on descriptions, not on identities
- Role-based access: operators can work without seeing faces
- Retention periods per data type, with automatic expiry
- Every search and export is logged
Safety and EHS, measured every minute
PPE, unattended objects, workstation activity, temperatures and standing safety checks. All from the same cameras, all with evidence attached.
PPE compliance by zone
Share of people seen wearing each item, today. Items are described in plain words, so a new item is one more word.
PPE breaches with evidence
Every person in the frame is checked. Select an item to focus on who is missing it.
Person crops filed with each breach
Illustrative sample frames, not from a customer site.
Unattended objects
A pallet that does not move makes no motion, so motion alarms never see it. EnlytiX checks marked areas on a timer and tracks how long each object has been left.
Dwell time for compliance
Aisles, fire lanes and secure bays must stay clear. You draw the region, name what to watch for in plain words (“a cart”, “a pallet”, “a person”), and set the rule: alert if an object sits unattended longer than 30 minutes. The clock pauses while someone is working with it and restarts when they leave.
How the rule works
- Region. Drawn once on a live frame (yellow). Objects outside it are ignored.
- What to watch. Objects and a person class in plain words; test them on a live frame before saving.
- Attended or not. A person beside the object pauses the clock, so a worker loading a cart never triggers it.
- Threshold. Unattended past the limit raises an alert with the frame, the time and how often it was attended.
What it is used for
- Fire lanes and emergency exits kept clear
- Aisles free of parked carts and trolleys
- Secure bays: nothing left without an owner
- Pallets and goods waiting too long to move
- Audit evidence that walkways stayed clear, by shift
Health is shown, not assumed
Each watch reports samples per hour, errors, and how long tracks live. “No alerts” is never ambiguous between “nothing was left” and “the camera could not see”.
Workstation activity
Activity index per station in 30-minute slots: 0 is the station’s own quiet level, 100 its normal working level. Shift start, break and end show up without any shift calendar.
Thermal readings
Maximum temperature per thermal camera, last 24 hours, with the alert line.
Standing safety checks
Yes/no questions asked on every relevant frame. A check that never fires is reported too, because silence is not proof the floor was clear.
Count stacks and objects from a camera
Cartons on a pallet, garment bundles on a table, fabric rolls on a rack: EnlytiX counts them from an ordinary camera frame, groups them into stacks, and reports how high each stack is. Pick the object type or click one example in the frame, and it finds the rest. No barcode, no scale, no training.
How it counts
- Choose what to count. Pick a type (carton, folded garment stack, poly-bag bundle, fabric roll, cut-panel bundle), or click one example item in the frame.
- Optionally draw a region. Only that rack, table or bay is counted.
- Every item is outlined. Items of the wrong size and duplicates are dropped automatically.
- Items are grouped into stacks and numbered from the bottom, so a person can check the count at a glance.
Where it is used
- Finished cartons ready for dispatch, counted at the loading bay
- Work-in-progress bundles at each line or table, per shift
- Fabric rolls on racks, checked against the stock record
- Stack height limits: a stack over the safe height is flagged
- Shipment evidence: a dated, counted frame for each pallet
Gates, perimeter and the camera estate
Who and what came in, how crowded it got, what the recorders heard, and whether every camera is actually working.
Draw lines and zones on any camera
Lines count crossings by direction; zones measure occupancy. Drawings scale with resolution, so they survive a camera upgrade.
Gate vehicles and number plates
Bangla and English plates, vehicle type, and whether the plate is on the approved list.
Occupancy against safe capacity
What the recorders report
Events from recorders and camera sensors are read the same way as video events.
Camera and recorder health
A camera that has stopped sending is a blind spot. Uptime, last frame and clock drift for every device.
Reports and notifications
The written work an operations team does every day, done for them: shift handovers, management briefs and buyer-audit evidence, in English or বাংলা, as Word or PDF.
Who gets told, and how
Route by severity and incident type. Each person sets their own threshold.
Patterns no single camera can show
Every incident links a camera, an area, a type, a time and the things named in its description. The graph connects them so you can ask questions that span cameras: what fires together, what makes an area different, which checks have never fired.
Select a node to see what it connects to.
Where EnlytiX earns its keep
The same platform, configured in plain language for each kind of site.
One platform, wherever a camera looks
EnlytiX needs no custom training. It reads each scene on its own and adapts to the place: a home, an estate, a factory, a warehouse, a campus, a street. What you want watched is written in a sentence, not trained into a model, so a new site or a new risk is minutes of work.
- Point it at your cameras. Any brand, existing recorders.
- Write what matters. One sentence per thing to watch for.
- Let it read the scene. It judges context itself, can answer “nothing is wrong”, and tells you why when it alerts.
What changes for your team
Fewer, better alerts
Every event is read before anyone is interrupted, and the system can say “nothing is wrong”. Operators see the few that matter.
Answers in seconds
Ask a question or describe a scene instead of scrubbing hours of recordings across many cameras.
Any site, no custom training
A new hazard or a new kind of site is a sentence, not a project. No training data, no model update, no downtime.
Evidence on demand
Audit packs, handovers and briefs with dated frames, ready when the buyer or inspector asks.
Your data stays home
On-premise and able to run fully air-gapped. Video and faces never leave your network.
Keep your cameras
Works with the cameras and recorders you have. No rip-and-replace.
How it is delivered
A server room appliance, your existing cameras, and a rollout measured in weeks.
Your site
IP cameras, any brand
Existing recorders (NVR)
Thermal and plate cameras
RTSP / ONVIF
on your LAN
on your LAN
On-premise appliance
Tier 1 Human Analyticsperception, faces, plates, counting
Tier 2 EnlytiXreasoning, search, agent, graph, reports
Event and evidence store
browser, app,
email, relays
email, relays
Your people
Command centre dashboard
Mobile app
Sirens, strobes, door relays
Open API to your systems
Hardware options
Huawei server platform
Huawei servers with AI accelerator cards
For organisations standardised on Huawei infrastructure or with sourcing requirements.
- Same software, same features
- Rack-mounted
- Sized on site survey
Edge box
Compact GPU appliance
For remote gates, small warehouses and sites with weak links. Reports up to a central EnlytiX.
- Fanless options
- A few to a few dozen cameras
- Works offline, syncs later
Multi-site
One command centre, many plants
Each site keeps its video locally. Only events and descriptions travel, which cuts bandwidth by over 95%.
- Location scoping per user
- Group-wide search and reports
- Central incident graph
Bringing your cameras on board
- Discover. Cameras and recorders on the network are found automatically (ONVIF), or added by address.
- Name and group. Each camera gets an area and a type: gate, warehouse, production, utility, outdoor, thermal.
- Draw. Lines and zones are drawn on a live frame in the browser.
- Watch live. Multi-camera live view with pan-tilt-zoom control, in any browser.
Sizing, as a starting point
Indicative only. Final sizing follows a site survey: resolution, frame rate and which features run on which cameras.
Rollout
- Week 1
Survey and connect
Site survey, appliance installed, cameras and recorders connected and grouped.
- Week 2
Describe the site
Incident types and checks written per camera type, with your EHS and security leads.
- Weeks 3–4
Pilot and tune
Alerts reviewed daily; thresholds and wording tuned on your own footage.
- Then
Expand
More cameras, more sites, more teams. No re-engineering.
Security and governance
- Runs entirely on your premises; can be fully air-gapped
- No video, face or event data leaves your network
- Role-based access with admin and viewer roles
- Audit trail of searches, exports and setting changes
- Open REST API with documentation for integration
See it on your own footage
A pilot runs on a handful of your cameras, on your premises, for four weeks.
- Emailinfo@sigmind.ai
- Webwww.sigmind.ai
- Productenlytix.live
- OfficeDhaka, Bangladesh
Briefing deck
Selected slides from the EnlytiX briefing. Select a slide to enlarge it; use the arrow keys to move through them.