Meet Edgar.

A compound Physical AI model built to understand the physical world.

Edgar combines specialized AI models, edge computing, and reasoning to understand what happens in physical spaces and what it means, in real time.

Changelog

Edgar 3.6CurrentDeeper checks and clearer diagnostics.
  • A second AI step can verify a finding using images, video or both.
  • One-click feedback on any event helps Edgar learn what matters to your business.
  • Smarter troubleshooting for cameras and video recorders, with suggested fixes.
Edgar 3.5PreviousPrivacy by design and camera tamper alerts.
  • Adjustable privacy levels blur faces or full bodies, set for each location.
  • Tamper detection sends an alert when a camera is blocked or moved.
  • Sales conversion heatmaps for stores connected to their point-of-sale system.
Edgar 3.4Sharper visitor counts and sales in context.
  • More accurate staff presence detection keeps staff out of visitor counts.
  • Sales insights in local currency, including payments and refunds.
  • Role-based views show each person the screens that matter for their job.
Earlier ModelsBefore Sep 2026Earlier releases and previous models
Aug3.3Jul3.2May3.1Apr3.0Previous Models

Release notes for models before Edgar 3.0 are shared on request.

Request more information

From perception to understanding.

Edgar follows the Edge Signal method: Detect, Understand, Improve, with each stage building on the last.

13:00 A store on a Saturday afternoonTime-lapse · 1 minute in half a second Example dataSimulation
DISPLAY ADISPLAY BFEATURE TABLEDISPLAY CSHELVINGWINDOWSTOCKROOMDESKENTRANCEQUEUECHECKOUTWALL BAYR1R2CAM 01CAM 02CAM 03
What is visiblePerception
In view
0
Moving
0
Still
0
Customers
0
Staff
0
At checkout
0
What is happeningEvents

No events yet. Observations are within the usual range.

    EdgarWhat it means

    Traffic is steady and checkout capacity matches demand. Nothing in the current observations departs from the expected early-afternoon pattern.

    A simulated afternoon, 13:00 to 14:30: arrivals rise, a queue forms while one associate is in the stockroom, a second register opens and the peak clears. Edgar's reads show the kind of reasoning described here; they are not recorded output.

    Read the technical detailHide the technical detailHow Edgar is built, level by level · how much each level keeps · why time changes the problem

    How Edgar is built

    Edgar is built as a stack of specialized models, each working at a different level of meaning. Perception finds the people and objects in every frame and follows them over time. Context places them in the space and on the clock: zones, paths, dwell and queues. Interpretation recognizes what is happening: a queue forming, a customer waiting, an area left unattended. Reasoning, drawing on language models, relates it to how the business runs and what is usual here, at this hour, with these staff.

    Each level hands the next something smaller and closer to meaning. The example below follows a store’s checkout on a busy Saturday, with the record Edgar keeps at each level.

    01Detect

    What is there, and where?

    Cameras already in the store capture the space as a stream of frames. Perception models find every person and object in each frame and follow them as tracks, so the customer at the entrance and the one at Display B a minute later are the same track. Nothing is interpreted yet: Detect reports what is there, and where.

    1. Input

      Physical environment

      A real space, sampled by cameras as a stream of frames.

      frame[t] · RGB · W×H · cam_03 · continuous
    2. TRACK 14

      Level 1Perception

      Perception

      People and objects found, located and followed as tracks.

      { id: 14, class: "person", bbox: [412, 233, 38, 96] }

    02Understand

    What is happening?

    Context turns positions into places and durations: track 14 has spent 4 minutes 12 seconds at Display B with no staff nearby, and five tracks have stopped in a line at checkout. Interpretation turns those patterns into events, a wait without help and a forming queue, and each event can be traced back to the tracks that produced it.

    1. 4:12DISPLAY BCHECKOUT

      Level 2Context

      Spatial and temporal context

      Positions become zones, paths and durations.

      { id: 14, zone: "display_b", dwell: "4m 12s" }
    2. UNASSISTEDQUEUE · 5

      Level 3Interpretation

      Interpretation

      Patterns across tracks become events.

      { event: "queue_forming", zone: "checkout", n: 5 }

    03Improve

    What does it mean, and what should happen?

    Reasoning reads the events in context: the hour, this store’s usual Saturday and where staff are right now. Edgar concludes that floor coverage is low just as demand peaks. An AI Agent turns that into one action for the person who can take it, and the result is measured, so what worked becomes the baseline for next time.

    1. Insight

      See what it means

      Edgar reads the events in context: the hour, this store's usual Saturday and where staff are right now.

    2. Action

      Act at the right moment

      An AI Agent turns the insight into one clear action and sends it to the person who can take it.

    3. Outcome

      Measure and learn

      The result is checked against what was expected, and what worked becomes the baseline for next time.

    One moment in a store, detected, understood and improved. Field names and values are examples. Example data

    How much each stage keeps

    Illustrative

    A video frame holds millions of values. By the time Edgar reasons about the moment, it fits in a sentence: each stage keeps only what the next one needs. Because the later stages work on compact records instead of video, they can hold weeks of history and compare many locations.

    Values needed to describe the checkout exampleLog scale

    One 1080p frame is 1920 × 1080 pixels × 3 color values. The later counts come from the records under each stage above, for twelve people and two events. Charcoal bars are Detect, blue bars Understand and Improve. Illustrative

    Why time changes the problem

    Conceptual

    Spatial intelligence answers where: where people and things are, what they are near and how the space is used. Temporal intelligence answers when and for how long: in what order things happened, how quickly a condition developed and whether it is clearing.

    Most conditions that matter to a business are both at once. A queue is a spatial arrangement that lasts. A wait without help is a duration defined by the absence of a nearby staff member. A peak is a rate measured against history. None of them can be read from a single frame.

    One frame, or the minutes around itCheckout at 13:47
    One frameCheckout5 people near the counter.Queueing, or just nearby?Over timeUsual for this hour on a Saturday · 2Growing: 1 → 5 in 12 minutes13:470513:3513:4013:4513:50People in line at checkout

    A single frame shows where people are. Only the minutes around it show that a line is forming, how fast, and that it is unusual for the hour. Example data

    Engineered for the physical world.

    Real stores are harder than curated datasets. These nine engineering problems shape Edgar.

    Environmental variability

    Light, crowds and layouts change; perception has to stay consistent and say when it is unsure.

    Daylight gives way to artificial light. Reflections, glare and shadows move. Merchandising changes how the space looks from week to week, and crowds hide people from view at exactly the moments that matter most. Perception has to give consistent results across all of this, and the layers above it need to know when an observation is less certain than usual, so the uncertainty is weighed instead of silently dropped.

    Even lightConfident detection
    Low lightReduced contrast
    OcclusionExtent partly inferred
    Oblique angleDistorted geometry

    The same observation under four conditions. Solid outlines mark confident detections; dashed outlines mark detections whose extent is partly inferred. Example data

    Privacy and data minimization

    Collect no more than the purpose needs, starting with processing imagery on site.

    Physical spaces contain people, and a system that observes them should collect no more than its purpose needs. Processing imagery close to where it is captured lets the rest of the system work with structured observations instead. Data minimization shapes what every layer computes, keeps and sends.

    Edgar is designed with privacy at its core. Imagery is processed on site, privacy masking such as face blurring is part of that processing, and data retention is configurable.

    Privacy masking, set for each locationConfigurable by location
    Faces blurredThe person stays visible
    Full bodies blurredOnly the outline of activity remains
    { id: 14,zone: "display_b",dwell: "4m 12s" }Structured recordWhat the layers above perception use

    Masking can be set for each location, and the reasoning layers work mainly from structured records like the one on the right. Example data

    Operational reliability

    The system notices its own failures and keeps working as well as it can.

    A model running in a store has no one watching it. It has to notice a camera that has been moved, blocked or knocked out of focus, a stalled stream or a site that has gone offline, and keep working as well as it can. Because perception runs on site, many operations can carry on locally when the connection to the cloud drops.

    A site goes offline, on-site processing carries onOne afternoon
    ConnectionOfflineObservationsProcessed on siteCloud reasoning resumes

    How edge processing can carry on through a dropped connection. Charcoal squares are observations made on site while offline. Conceptual

    Continuous observation

    Streams never end, so state has to stay bounded, revisable and recoverable.

    A model that answers one question and stops can treat each input on its own. Edgar keeps a running account of each environment, updated as new frames arrive. Memory has to stay bounded while history grows, identities have to persist through gaps in coverage, and small errors must not build up into drift over hours of operation. State has to be compact, revisable when later evidence contradicts it, and recoverable after an interruption.

    History grows; the state kept stays boundedOne store, one day
    Frames observed · keeps growingState kept · bounded, compacted as it goes09:0011:0013:0015:0017:00

    Shapes are schematic, not measurements. Conceptual

    Real-time processing

    Perception keeps pace with the stream; deeper reasoning runs when it adds value.

    Some interpretations are only useful while the situation is still unfolding. Perception and tracking run on a continuous clock, frame by frame, on site. Interpretation and reasoning run when events call for them, so they can take more time per decision.

    Real-time path · on site
    Capture
    Decode
    Perception
    Tracking
    Event
    Interpretive path · typically cloud
    Transmit
    Orchestrate
    Reason
    Output

    The main steps on each path. Not every observation takes every step, and where each step runs can vary by deployment. Conceptual

    Scaling across locations

    Every site has its own context, so observations must mean the same thing everywhere.

    Each new location brings its own cameras and systems, compute, floor plan and context. Scaling needs configuration that is defined once and applied consistently, visibility into every site, and models that behave predictably in places they were not tuned for. Two sites can only be compared if a zone, a dwell or a queue is defined the same way in both.

    One definition, every layoutThree example sites
    StoreWide sales floor · 4 in the queue
    Bank branchL-shaped hall · 2 in the queue
    CaféNarrow counter · 5 in the queue
    QueueDefined once, then applied to each floor plan, so a queue of 4 means the same thing at every site.

    The dashed blue area is the queue zone each site maps to the same definition. Example data

    Different cameras and hardware

    Varied cameras and compute are evened out so results compare across sites.

    Cameras differ in resolution, frame rate, lens, compression and mounting height, and on-site compute differs from one location to the next. Perception has to even out these differences, so that a duration measured on one camera means the same as one measured on another.

    Edgar runs on Edge Signal's hardware-agnostic edge architecture and works with compatible camera systems already in place. Compatibility is reviewed as part of each deployment.

    Three cameras, one measurementThe same customer at Display B
    4m 12sWide-angle lensHigh on the ceiling, bowed edges
    4m 12sLower resolutionFewer pixels per person
    4m 12sLow, angled mountClose and in perspective
    Dwell4m 12s on all three, so the stores they sit in can be compared.

    Three example cameras watching the same display. Perception is designed to even out lens, resolution and mounting height before anything is measured. Example data

    Model updates and deployment

    Updates are managed centrally and can be staged across a network of locations.

    An update has to reach many sites with different hardware and be checked against real environments before it takes full effect. In a compound system, components can be updated on their own: a perception model can improve without retraining the reasoning layer, and the reasoning layer can change without touching a camera.

    How a model update can be staged across a network of locations, with a check between stages. Each square is one site. Illustrative

    Balancing compute with reasoning

    Computation is spent where it is needed, instead of on every frame.

    Inexpensive, continuous processing filters the stream. More expensive interpretation runs on the share of observations that call for it. The most capable reasoning runs when the context or a question requires it, so the cost of the system follows what is happening in the space more than the number of frames recorded.

    As data moves up the stack its volume falls and the cost of handling each item rises. Heights are schematic. Conceptual

    Compound AI on distributed infrastructure.

    Many specialized models working as one, across Edge AI on site and Cloud AI with LLMs.

    Cameras
    Already in place
    Stream ingest
    Decode, normalize, timestamp
    Visual perception models
    Detection, localization, attributes
    Tracking and local state
    Identity, position, zone, dwell
    Event candidates
    Structured observations, local events
    Unified intelligence
    One consistent view of each location
    Reasoning models
    Interpretation, relationships, language (LLMs)
    Context and memory
    Site history, baselines, cross-site state
    Orchestration
    Routing, scheduling, context passing
    Your teams
    Alerts, answers, follow-ups
    Your systems
    Workflows and data
    In the store
    On site · Edge AI
    Cloud AI · LLMs
    Your business
    Site boundary
    Raw video
    Processed on site
    Model updates
    and configuration
    Structured observations
    and selected masked imagery
    Local inference · real time · efficiency-bound
    Many locations · long history · larger models
    Other systems
    Operational data and sensors
    Remote updatesHealth alertsTamper alertsPrivacy by location

    Edgar's infrastructure at a conceptual level: the cameras and systems a location already has, Edge AI on site, and Cloud AI with LLMs. Work flows down through the site, across the site boundary and up through the cloud; the moving squares show what crosses it. In the standard architecture, live video is processed on site; structured observations, and selected privacy-masked imagery where used, go to the cloud for deeper analysis.

    Read the technical detailHide the technical detailFour levels and the ideas behind them · what makes a model compound · where each level runs · on-site infrastructure · compared with one general-purpose model · how Edgar uses language models · how context builds up

    Four levels of the stack

    Conceptual

    Edgar’s work is organized in four levels. Each answers one question and builds on the one before it, and the ideas behind each level are listed with it.

    1. Level 1
      PerceptionWhat is present?

      Detect, locate and track people and objects in each camera view, frame by frame. Literal by design: it reports what is in view without deciding what it means.

      The ideas behind it

      Computer vision and visual perception. Pixels become a literal description of the scene, which keeps the foundation fast, testable and reusable across very different places.

      Object detection and tracking. Detections are linked across frames into identities that survive occlusion, crowds and people leaving and coming back into view.

    2. Level 2
      ContextWhere, and for how long?

      Place observations in a shared map of the space and a continuous timeline. Keep state across camera views: identities, zones, paths and durations.

      The ideas behind it

      Spatial relationships. Positions in the image become positions in the space: inside a zone, beside a fixture, a few meters from another person.

      Temporal understanding. Durations, sequences and rates exist only across many observations: seconds for movement, minutes for activities, weeks for patterns.

    3. Level 3
      InterpretationWhat is happening?

      Recognize events, activities and conditions from patterns in tracked behavior: a queue forming, a wait without assistance, an area left unattended.

      The ideas behind it

      Event and activity recognition. Every event is defined by the tracks and zones underneath it, so it can be traced back to the observations that produced it.

    4. Level 4
      ReasoningWhat does it mean?

      Relate what is happening to history, to other locations and to how the business runs. Explain, compare and answer questions in plain language.

      The ideas behind it

      Contextual reasoning. The same observation means different things in different settings: five people at a counter is routine at lunchtime and worth a look at an unusual hour.

      Relationships between observations. Relating observations across zones, cameras and people is where a compound system adds most to what any single detector could report.

      Patterns over time. Repeated observations become the baseline the present is judged against, so Edgar can say something is unusual, not only that it happened.

    Four levels of the compound stack. Each builds on the one before it; where each one runs is shown below. In the Edge Signal method, Detect is Level 1, Understand covers Levels 2 and 3, and Improve builds on Level 4.

    What makes a model compound

    Conceptual

    Many AI systems are organized around a single model: an input goes in, an output comes out, and every capability lives inside one set of weights. A compound AI system splits the task across several components, each specialized for part of the problem, and coordinates them so that the output of one becomes the input or the context of another. Its intelligence comes from the whole arrangement: the components, the information that flows between them, and the logic that decides what runs when.

    Physical environments make that split necessary. Understanding a space takes dense visual perception at video rate, a persistent account of where things are and how they move, memory measured in hours and weeks, and open-ended reasoning that connects an observation to what it means for the business. Perception has to run continuously, close to the camera, on every frame. Reasoning benefits from broad context and from large models that would be impractical to run at that rate. One model asked to do both is either too slow to perceive or too narrow to reason.

    Each part of a compound system can be specialized, inspected and improved on its own. Conceptual

    Where each level runs

    Conceptual

    Perception models run on site, beside the cameras. They are dense, fast and literal: they find and follow people and objects in every frame without deciding what it means. Reasoning models, including language models, typically run in Cloud AI, where they can hold context across weeks of history and every location. Context and interpretation sit in between and can run in either place, and supported on-premises configurations can bring more of the stack on site. Edgar is the combination of all of them: one consistent view of each environment.

    WhereLevel 1PerceptionLevel 2ContextLevel 3InterpretationLevel 4Reasoning
    On site · Edge AIRuns hereRuns hereRuns hereNot here
    Cloud AI · LLMsNot hereRuns hereRuns hereRuns here

    The general tendency of each level; the split can vary by deployment, and not every capability can run in every environment.

    The timescales each layer works onLog scale, from a single frame to months
    Edge AI · frames to minutesCloud AI · minutes to months10 ms1 s1 min1 hour1 day1 week1 montha framea 4-minute dwella Saturday baselineFrames at 30 per second sit at the far left; baselines built over weeks at the right.

    Edge AI works on frames to minutes; Cloud AI on minutes to months. Conceptual

    On-site infrastructure

    Edgar runs on Edge AI hardware on site, alongside the cameras and systems a location already has, and is managed remotely.

    Your infrastructure

    Works with compatible cameras already in place, on Edge Signal's hardware-agnostic edge architecture.

    On-site processing

    The Edge AI device analyzes video in the location, close to where it is captured.

    Remote updates

    Updates and restarts are managed remotely, across every location.

    Health alerts

    An alert when a camera or device goes offline, with guidance on how to fix it.

    Tamper alerts

    An alert when a camera is blocked or moved.

    Privacy by location

    Privacy masking, such as face or full-body blurring, set for each location.

    Why a compound architecture

    Conceptual

    A compound design matches each kind of computation to the conditions it needs. Fast, efficient components keep pace with the world. Slower, more capable components get the context and the time to think. The steps in between become explicit records that can be inspected, compared and improved one at a time. The comparison below describes properties of each design, not benchmark results.

    PropertyOne general-purpose modelCompound system
    LatencyEvery output waits on the largest model in the path.Fast components run continuously; slower reasoning runs when it adds value.
    SpecializationOne set of weights is a compromise across many tasks.Each component is trained, tuned or chosen for a narrow job.
    Memory of timeLimited to what fits in the model's input at once.State and memory persist across hours, days and locations.
    Data movementRaw inputs travel to wherever the model runs.Compact, structured observations can travel in place of raw video.
    InspectionThe reasoning in between stays hidden inside the model.Detections, tracks and events are explicit and can be audited.
    ImprovementImproving one capability means changing the whole model.Components can be improved, tested and replaced one at a time.
    ComputeThe same cost is paid for every input.Expensive computation is saved for the observations that need it.

    How Edgar uses language models

    Conceptual

    Language models take part in the system, and they reason over evidence that other components have already extracted, organized and placed in context: tracks, zones, durations, sequences and their history.

    The answer is built from the evidence above. Example data

    Compound orchestration

    Conceptual

    Orchestration is what makes the two layers one model. It decides which component handles an observation, carries context from one component to the next, and combines their outputs into one consistent interpretation. Each output keeps a record of where it came from, so a conclusion can be traced back to the observations behind it.

    1. Edge AIPerception
      track 14 · person · cam_03 · 13:47:05
    2. Edge AILocal state
      track 14 + zone display_b · still for 4m 12s · no staff nearby
    3. Cloud AIContext
      … + usual dwell here at this hour 1m 50s · 2 similar events in 30 min
    4. Cloud AIReasoning
      Edgar

      A long wait without help at Display B, unusual for this time of day and during a checkout peak.

    How context builds up around one observation between the site and the cloud. Example data

    See what Edgar understands in your spaces.

    Bring a location, and we’ll show you what Edgar detects, understands and helps improve there, with the infrastructure you already have.