A Softmachines work · Video and sensor intelligence
Argus gives a team a way to turn selected camera and sensor signals into evidence it can inspect and act on while there is still time to change the outcome.
The premise
A site does not need another archive of footage, nor an alarm for every pixel that changes. It needs a defensible way to state what matters, see whether the evidence supports it, and respond before a moment turns into a problem.
Argus keeps that path visible. It starts with the site and its coverage, asks a model only when a meaningful change earns a closer look, and makes a candidate clear a policy and corroboration path before it can ask for a person's attention.
A signal earns attention
A site is not an abstract set of feeds. It is a set of cameras with names, locations, current status, and an operator who can check whether the picture is still available. Argus begins there, so a missing camera is never silently treated as coverage.
The coverage view also makes the next question concrete: which camera is responsible for a condition, and which part of the site does that condition describe?
The live view keeps current frames, camera names, and availability together. Before a rule, model, or incident asks for attention, an operator can tell the difference between a quiet scene and a missing signal.
That makes coverage a living operational map. The condition has a specific world to watch, and the team can see whether that world is still present.
A rule begins with plain language, then gives that condition an accountable scope: the site, an individual camera or all available coverage, an optional monitoring zone, severity, and whether it starts in dry run. The condition is saved as written, without turning incidental visual details into a requirement.
Before it can act, the rule can be tested and bounded by a cooldown. The people responsible decide what is worth watching, where it applies, and how cautious the system should be before a condition becomes operational.
The inbox keeps new, acknowledged, investigated, resolved, and false-positive events in separate but connected states. An operator can filter the work, open the evidence, and make an explicit outcome instead of letting an alert disappear into a wall of notifications.
It does not pretend that every candidate is true. Its value is that the people responsible can see what still awaits a decision and can trace what was already decided.
An incident holds the written condition, site, camera, timestamp, captured evidence, confidence, and current resolution state in one view. The operator does not have to reconstruct why a frame was surfaced from a separate system.
That matters because an image alone is not a conclusion. Argus frames the evidence as evidence, then records the human decision that follows it.
A single incident can be one part of recurring activity or a multi-camera event. Cases keep those incidents together under a named owner, priority, and operational context, without overwriting the evidence in the original record.
The result is a real handoff surface rather than a queue someone must hold together from memory.
A model with boundaries
Argus is selective before it is intelligent. An edge gate makes a quiet camera free to ignore; a second, block-aware check lets a local change earn a closer look. Only then does a vision model assess the site's written condition, and its candidate enters a deterministic corroboration path that can confirm, clear, or keep gathering evidence.
Gate twice
Start at the edge, then preserve a meaningful local change at the control plane before a model call.Corroborate over time
Accumulate supporting and counter-evidence from distinct observations, with stale evidence losing weight.Decide with a person
Surface the condition, evidence, and resolution together so someone accountable can decide what follows.
The analysis history records the model response, deterministic rule gate, and each observation's place in the corroboration path. It shows whether the system is still gathering evidence, has cleared it, or has opened an incident. A confident negative counts too, so the record can show why an early candidate was not promoted.
This is the practical difference between a confidence score and a decision: the evidence that changed the state remains available to inspect.
Each organisation can configure, test, and identify its own model connections. A site can use a chosen approved endpoint or an organisation default, while endpoint and model details stay visible to the people who operate it.
The platform is not committed to one provider. It is committed to the same inspectable policy and corroboration path whichever approved model a site uses.
Operational ground
Argus connects the edge through a private overlay rather than exposing RTSP feeds or credentials to the public internet. The administration surface makes the service endpoint, workspace roles, and site connection legible to the people who run it.
That separation is practical. A quiet camera, a disconnected gateway, and a missing permission are different conditions. Argus records them as different conditions before an operator is asked to interpret video.
More than alerts
Argus uses the same discipline for reports and scene narratives. Each account is built from stored observations, incidents, and coverage data, so a shift can be understood without turning missing footage into a claim.
The reports view deliberately selects cameras and a time window. It makes clear that a narrative uses the stored scene notes produced by those cameras, rather than claiming to see footage that was never gathered.
A daily or shift report becomes a reviewable operational artifact, not a generic AI summary.
Here Argus presents a camera narrative with timestamped notes and a bounded summary of what stayed consistent and what changed. This is not an incident screen. It is the useful story between incidents, with enough detail for a person to decide whether the ordinary pattern deserves a closer look.
The account remains reversible: it can be sent on demand, scheduled for delivery, or deleted when it no longer belongs in the record.
Daily reports and scene narratives can be scheduled in the site's own local time. A rule can also observe only during its relevant arm window and rest during a cooldown, so Argus is designed for the second and hundredth occurrence, not only the first demonstration.
Work that keeps moving
Argus treats recordings and notifications as operational work, not as loose files and one-shot messages. The record remains observable from the source capture through the workflow and the delivery outcome.
Workflow stages
ThumbnailTimeline spriteObject classificationCount / occupancyNotificationCustom serviceProcess
A site or camera can route its recordings through a versioned workflow. Each run has its own progress and result, so it can be observed or retried without having to re-trigger the original recording.
Deliver
Argus can deliver evidence through Telegram, WhatsApp, or a customer webhook. The actual image or short GIF reaches the messaging service, while a webhook can receive the link it is able to access.
Recover
Delivery retries respect provider backoff. A message that still cannot be delivered becomes a visible dead letter an operator can inspect and retry, instead of a failure hidden behind a green dashboard.
Aggregate retail insights
Retail Insights does not feed the incident pipeline. It has its own privacy boundary, policy, model-release gate, reporting surface, retention, and audit trail. It is designed for aggregate movement and apparel trends, not recognition, re-identification, or individual tracking.
The reporting view is designed to answer operational questions without making a visitor the unit of analysis: when a site becomes busier, how movement shifts by zone and direction, which weekday or campaign pattern has changed, and whether the reading is supported by enough healthy coverage to use.
Argus retail demo
22 Sep – 18 Oct 2026 · 27 days of aggregate buckets from the Argus demo workspace · looped demo footage, no individual is tracked
Camera detections, not unique shoppers. UTC days.
| Date | Detections |
|---|---|
| 22 Sep | 648 |
| 23 Sep | 1,107 |
| 24 Sep | 278 |
| 25 Sep | 4,733 |
| 26 Sep | 9,740 |
| 27 Sep | 7,662 |
| 28 Sep | 5,119 |
| 29 Sep | 2,665 |
| 30 Sep | 4,296 |
| 1 Oct | 7,086 |
| 2 Oct | 7,879 |
| 3 Oct | 6,215 |
| 4 Oct | 6,616 |
| 5 Oct | 8,683 |
| 6 Oct | 8,929 |
| 7 Oct | 6,759 |
| 8 Oct | 6,573 |
| 9 Oct | 8,263 |
| 10 Oct | 11,140 |
| 11 Oct | 11,988 |
| 12 Oct | 10,050 |
| 13 Oct | 4,657 |
| 14 Oct | 8,170 |
| 15 Oct | 7,641 |
| 16 Oct | 8,102 |
| 17 Oct | 3,649 |
| 18 Oct | 89 |
The seven-day average rises from 5,900 to 8,821 a day between 1 and 16 October. The first three and last two days are capture ramp-up and wind-down; they are shown, but kept out of every comparison.
Calendar heatmap, one cell per day
Weekends carry the strongest cells; the second weekend of October is the high point of the report.
Complete Fri–Thu weeks
| Week | Detections |
|---|---|
| 25 Sep – 1 Oct | 41,301 |
| 2 – 8 Oct | 51,654 |
| 9 – 15 Oct | 61,909 |
Each week is compared with the one before, on the same weekdays, never with a partial one.
Detections by UTC hour, whole period
| Hour (UTC) | Detections |
|---|---|
| 00:00 | 3,320 |
| 01:00 | 3,859 |
| 02:00 | 3,266 |
| 03:00 | 4,895 |
| 04:00 | 4,817 |
| 05:00 | 5,224 |
| 06:00 | 5,475 |
| 07:00 | 5,292 |
| 08:00 | 10,478 |
| 09:00 | 11,974 |
| 10:00 | 13,092 |
| 11:00 | 10,738 |
| 12:00 | 13,835 |
| 13:00 | 16,099 |
| 14:00 | 13,011 |
| 15:00 | 8,855 |
| 16:00 | 6,850 |
| 17:00 | 5,424 |
| 18:00 | 4,491 |
| 19:00 | 4,961 |
| 20:00 | 5,218 |
| 21:00 | 2,499 |
| 22:00 | 2,412 |
| 23:00 | 2,652 |
Average detections per weekday, 22 complete days
| Weekday | Detections |
|---|---|
| Mon | 7,951 |
| Tue | 5,417 |
| Wed | 6,408 |
| Thu | 7,100 |
| Fri | 7,244 |
| Sat | 9,032 |
| Sun | 8,755 |
Saturday and Sunday lead; Tuesday is the quietest weekday.
Share of 168,700 attributed detections
Sleeve and leg length, combined
Share of detections where the cue was present
5,071 buckets by privacy outcome
Default minimum cohort: 20 people before an aggregate is reported; smaller groups stay withheld. Apparel cues only, no faces and no identities.
Mean detection confidence by week
Coverage comes from one street camera (99.8% of detections) plus two near-silent ones. The report says so, so a change in coverage is never read as a change in demand.
Occupancy heatmap, 32 × 18 grid over the camera view
| Rank | Column | Row | Samples |
|---|---|---|---|
| 1 | 23 | 10 | 24,401 |
| 2 | 27 | 11 | 22,306 |
| 3 | 23 | 11 | 21,647 |
| 4 | 11 | 11 | 21,523 |
| 5 | 24 | 10 | 21,382 |
| 6 | 26 | 11 | 20,805 |
| 7 | 24 | 11 | 20,427 |
| 8 | 25 | 10 | 20,279 |
| 9 | 25 | 11 | 19,781 |
| 10 | 25 | 9 | 18,475 |
Occupancy gathers in the lower half of the view: 218 cells ever used, and the busiest 10% of them hold 28% of all samples. Only cell counts leave the camera, never a track.
Net direction between neighbouring cells, 16 × 9 grid
| Rank | From column | From row | To column | To row | Net count |
|---|---|---|---|---|---|
| 1 | 11 | 5 | 12 | 4 | 1,248 |
| 2 | 13 | 5 | 13 | 6 | 1,211 |
| 3 | 12 | 5 | 12 | 4 | 965 |
| 4 | 11 | 6 | 12 | 5 | 924 |
| 5 | 10 | 6 | 11 | 5 | 875 |
| 6 | 11 | 5 | 12 | 5 | 813 |
| 7 | 12 | 4 | 13 | 4 | 800 |
| 8 | 10 | 6 | 12 | 5 | 766 |
| 9 | 12 | 5 | 13 | 4 | 754 |
| 10 | 9 | 6 | 11 | 5 | 688 |
Each arrow is the net flow between two cells after the opposite direction is subtracted, so back-and-forth jitter cancels and only a lasting current remains. Thickness follows volume.
Same grid, split by UTC time of day, each scaled to its own peak
00–07 UTC22% of samples
08–15 UTC65% of samples
16–23 UTC12% of samples
Zone-to-zone walks on an 8 × 5 grid
Next seven days, from the same weekday in the last three complete weeks
| Weekday | Min | Mean | Max |
|---|---|---|---|
| Sat | 6,215 | 9,032 | 11,140 |
| Sun | 6,616 | 8,755 | 11,988 |
| Mon | 5,119 | 7,951 | 10,050 |
| Tue | 2,665 | 5,417 | 8,929 |
| Wed | 4,296 | 6,408 | 8,170 |
| Thu | 6,573 | 7,100 | 7,641 |
| Fri | 4,733 | 6,958 | 8,263 |
A plain same-weekday band for this page. In the product, the forecast shows only after complete comparable days, with its back-tested error and the “same as last week” guess it has to beat.
Daily traffic, local hours, directional flow, and the camera contribution behind them make staffing, replenishment, and service placement discussable as a question about a place and a period.
Period, zone, display attention, and apparel context can be held against a selected baseline. Supported movement and an early signal have different states, and both remain visible.
Capture gaps, withheld cohorts, changed models, and weak agreement are part of the same reading. That makes the next check specific without pretending that an aggregate trend proves cause or conversion.
Aggregate only
A temporary, per-camera track prevents duplicate counting only while someone crosses a zone. Once an aggregate bucket is written, its track, crop, and reusable embedding are discarded.
Suppress by design
A cohort below the configured minimum is suppressed before storage. Sparse attribute combinations are redacted too, so a trend never becomes a back door to an individual.
Earn the report
A site cannot enable reporting until its model release passes fixed benchmark checks. Data-quality signals stay beside every trend, so a quiet floor and an unreliable camera are never presented as the same thing.
The value is a clearer picture of a place, without making any visitor the subject of that picture.
Argus keeps site access private through an overlay network, so cameras do not need to be exposed to the public internet. Edge health, queue pressure, retention, and audit history remain part of the same operational surface as the incident itself.
Organizations, sites, memberships, and roles shape every action. The system can coordinate a large estate without making the local owner, the camera, or the original condition disappear from the record.
That is the real value of attention here: less time scanning what is ordinary, more time understanding the moment that changes what someone should do next.
Argus gives people more time to notice, and more evidence to decide.
Return to works