Product · Caddy AI

Caddy AI — AI video surveillance and practical machine learning on your own data

Computer vision for the cameras you already own and language models that work with your documents, deployed on-site, in-country or in the cloud depending on your rules.

Caddy AI brings AI video surveillance and practical machine learning to organisations in Tashkent and across the region. Instead of adding more screens to a control room, it watches the feeds for the events you care about and sends a short, verifiable alert when one happens.

The same platform also hosts language models for internal documents: procedures, contracts and technical manuals. Staff ask questions in English, Russian or Uzbek and receive answers with links to the source pages, while the documents stay where your policy requires.

We keep the scope narrow and measurable. Each project begins with a pilot on your own footage or files, and the results — including the false alarms — are shared before anything is rolled out. Use cases that do not reach the agreed quality are not deployed.

Caddy AI usually runs on Caddy Edge nodes next to the cameras, with results flowing into Caddy Analytics dashboards. When cameras or cabling need work first, our CCTV and access control team handles the installation, and the automation and AI integration team connects alerts to your existing systems.

The problem

Cameras record everything and nobody watches

  • Operators cannot watch every screen

    A control room with forty camera feeds and two operators misses most events. Footage is usually reviewed only after something has already gone wrong.
  • Generic AI tools do not fit local reality

    Off-the-shelf models are trained on other countries, other uniforms and other lighting. Accuracy drops on your sites and false alarms erode trust.
  • Sensitive data cannot leave the building

    Video of employees, patients or citizens, and internal documents, often must not be sent to foreign AI services.

The solution

AI video surveillance tuned to your sites

Caddy AI runs detection and search models close to your cameras and data, and is tuned on footage from your own environment during a pilot.
  • Event detection

    Intrusion into restricted zones, missing helmets or vests, loitering, smoke and vehicles in the wrong place, with alerts to the right people.
  • Occupancy and flow

    Counts of people and vehicles per zone and per hour for stores, entrances, canteens and loading bays.
  • Video search

    Find footage by time, zone, object type or colour instead of scrolling through hours of recordings.
  • Document assistants

    Language models that answer questions from your contracts, manuals and procedures, hosted where your policy requires.
  • Human in the loop

    Every alert can be confirmed or dismissed, and that feedback is used to reduce false positives over time.

How we deliver

How we introduce AI without disrupting operations

  1. 01

    Use-case workshop

    We agree on two or three measurable use cases and the cameras or data sources involved.

  2. 02

    Pilot on your footage

    Models are configured and evaluated on recordings from your sites, and the results are shared openly.

  3. 03

    Deploy

    Approved use cases go live on edge nodes, in Caddy Cloud or in your data centre, connected to your alerting tools.

  4. 04

    Review and tune

    Monthly reviews of alert quality and adjustments to zones, schedules and thresholds.

What you can check

How you can judge the results

AI quality should be measured, not claimed. Each deployment includes evidence you can check yourself.
  • Pilot report with precision and recall measured on your own footage
  • List of confirmed and dismissed alerts available for audit
  • Data-processing description stating where video and documents are stored
  • Configuration of zones, schedules and thresholds per camera
  • Retention policy for clips and model logs agreed in writing

Technologies we work with

  • PyTorch
  • ONNX
  • NVIDIA DeepStream
  • OpenCV
  • ONVIF
  • RTSP
  • PostgreSQL
  • vector search
  • open-weight LLMs

FAQ

Frequently asked questions

Do we need to replace our cameras?

Usually not. Caddy AI works with ONVIF and RTSP cameras from most manufacturers. Very old or low-resolution cameras may limit some use cases, which the pilot will show.

Does video leave our premises?

Only if you choose so. Models can run on Caddy Edge nodes on site or in our in-country cloud, and the data-processing description documents the flow.

Do you use face recognition?

We do not enable face recognition by default. Any identification use case requires a legal basis, a documented assessment and your written approval.

How accurate is it?

Accuracy depends on camera placement, lighting and the event type. That is why every project starts with a pilot that measures results on your footage before a wider rollout.

Can the alerts go to Telegram?

Yes. Alerts can be sent to Telegram, email, SMS gateways or your existing video management and ticketing systems.

Can we use Caddy AI on documents only?

Yes. Document assistants are a separate module and can be deployed without any video use cases.

Talk to an engineer about your project

Tell us what you run today and what has to change. We reply within one business day with next steps, not a sales script.

Call usBook a consultation