| Prepared for | Gallery Condominiums Board |
|---|---|
| By | Elberta Labs |
| Date | 25 August 2026 |
| Re | AI-powered managed security partnership and 60-day pilot |
The building already runs IntuVision analytics. Its value today is limited by how noisy it can be and how effectively it reaches the front desk. The current integrator does not perform the recurring improvements needed to keep it tuned. Elberta has already resolved several of these at no charge, with Aref working directly with IntuVision as a resident of the building.
We looked for an off-the-shelf product to add smarter GenAI analytics on top of the cameras and IntuVision the association already owns. There isn't one. The options on the market require new hardware and a new platform, and they bill per camera per year with limited customization. Most offerings are catered to enterprise needs. The practical path is to build a layer on top of what is already in place, tailored to the building, on a foundation of open-source systems.
Although Elberta ran a dog-park proof of concept at no charge to confirm AI monitoring on the existing cameras works without per-license fees, the engagement model proposed is an ongoing operating role: an AI-powered managed security partner accountable for making the building's cameras, analytics, access control, and front-desk workflow function together, and keeping them working. This way the board buys a managed result: one accountable relationship for the cameras, analytics, access control, and workflow together.
What this covers is broader than any single thing the building pays for today. Elberta becomes the single point of coordination across the analytics, the reception workflow, and the camera and access-control vendors, work that today is scattered across separate relationships and unpredictable labor bills. The association keeps its own direct vendor contracts. What becomes one predictable number is Elberta's coordination and engineering, in place of scattered hourly and markup charges.
Elberta proposes to serve as the board's AI-powered managed security partner: the operating layer that makes the building's existing cameras, analytics, access-control vendors, and front-desk workflow work together reliably, with AI screening select cameras to surface the events the desk should see first.
Our engagement follows a simple arc:
What this is. A coordinating operating layer accountable for a defined scope: the analytics and reception-alerting pipeline, its health, tuning, and patching, plus coordination of the building's camera, access-control, and networking vendors when issues cross systems.
What this is not. Not a per-camera license. It does not replace the IntuVision license today; any future change would follow a documented evaluation against the alternatives. The system is an operator efficiency tool that assists the human front desk. It is not an autonomous security system, a life-safety or guard service, or a guarantee of absolute security, and liability for physical security at the property is not unlimited. The scope is limited to AI analytics on top of select camera feeds. Elberta's total liability under the managed service is limited to the fees paid, and insurance coverage is detailed under Continuity and independence below.
Cameras and analytics are only useful if the front desk sees the right events and acts on them. Today most of what the systems detect never reaches the desk in a usable form. The existing software is slow to load, cannot be read at a glance, and does not learn as reception marks alerts as false alarms. It runs on older computer-vision analytics: real, but not the generative AI that can reason about a scene. The reception software suite closes that gap. It is the product the desk actually uses: one dashboard that turns scattered detections into a workflow staff can follow.
Existing tools stay where they are useful. This integrates the building's systems, it does not rip and replace them. Reception adoption is an explicit success measure: the suite only works if the desk uses it daily.
The review runs a two-stage AI pipeline on the building's own cameras. Generative AI vision models understand what is happening in a scene, beyond the fact that something moved. They tell "a resident walked through the gate" apart from "a person is loitering at the gate."
Hybrid AI detection. The pipeline is deliberately hybrid. Everything that can run locally runs locally: all real-time detection runs on the building's own GPU, so the constant, high-volume work never leaves the property. The occasional second-stage scene review runs wherever it runs best. Where the on-prem GPU can carry it, it stays on-prem via Ollama. Where a larger hosted model is the better choice, the single flagged snapshot is sent over an encrypted connection to an enterprise API endpoint operating under zero-data-retention terms, meaning the image is not stored and is never used to train a model. Approved providers are limited to Anthropic, OpenAI, and Google Vertex AI on their enterprise data terms. Where a routing layer is used, it runs in zero-retention mode with request logging disabled. Resident imagery is never sold, retained, or used for training. Over time, as open models improve, more of this runs locally.
The safety guarantee is built into the pipeline. The AI can only downgrade an alert, never delete one, enforcing the downgrade, never discard rule described below.
The proof of concept shows the AI-assisted review reduces false-alarm noise without hiding real events.
On one gate camera (the Peachtree pedestrian gate) over a four-day sample, the existing system produced 40 alerts, about 10 a day, and most were benign: maintenance crews, residents, and pedestrians passing through. When nearly every alert is a false alarm, staff stop looking, and a real event slips past.
The AI-assisted review adds a second, smarter check on every alert. It keeps the camera's motion and zone detection, confirms a person is present, then asks a vision model what that person is doing, based on behavior and context only.
The safety rule: downgrade, never discard. A benign-looking event is quieted to a review tier and still logged, but never deleted. Only clearly harmless activity is quieted; anything uncertain or concerning still raises a full alert. The system can turn the volume down, never off, so a real event is never suppressed.
| On your cameras | What happens today (IntuVision) | With the AI-assisted review |
|---|---|---|
| Grounds worker with a leaf blower | About 5 "loitering" and "intrusion" alarms in roughly six minutes of yard work | Recognized as routine maintenance and quieted to the review log instead of paging staff five times |
| Person sitting in a parked car | Flagged as "loitering" with no security value | Read as a person in a parked car, no approach to the gate, quieted |
| Resident walking through the gate | Every crossing triggers an alert, so normal foot traffic is much of the daily noise | A plain walk-through with no gate tampering is logged quietly, keeping the alert stream meaningful |
| Person lingering by a scooter near the gate (over a minute) | Ambiguous; the kind of event where certainty is not possible from one frame | Judged "likely benign" and downgraded to the quiet review tier, kept visible for a supervisor, not deleted |
The engagement starts with a pilot: a single fixed fee, roughly 60 days, and a defined exit.
Success is measured operationally, against a bar agreed up front:
A consolidation objective. The pilot also measures how many cameras the local box can carry and how closely its detection matches IntuVision. Proven out, this gives the board a line of sight to running more cameras on one system over time and reducing the separate per-camera analytics subscription. This is a goal to validate during the pilot, not a commitment made in advance. It reflects Elberta's role as an independent advocate for whatever best serves the building, on a flat fee that rewards simplifying the system because it does not grow with the hours billed.
The exit, stated up front. If the pilot does not meet the agreed criteria, Elberta recovers its hardware, the parties stop, and there is no ongoing obligation on either side.
If the pilot meets the criteria, the engagement continues as a managed service at $1,000 per month ($12,000 per year), on a 12-month initial managed-service term. This is a flat fee for maintaining the stack, incrementally adding cameras, improving the alerts and noise, and managing coordination across the camera-system vendors and the reception operating layer. It is not a per-camera license. Either side may end the managed service at any time with 60 days written notice, for any reason. The 12-month term sets the pricing basis; the no-cause exit means the association is never locked in.
Year one is a founding-client rate, held below market as Gallery is Elberta's first deployment. At each renewal the rate is reassessed openly with the board. The association is never locked in: if a future rate cannot be agreed, the 60-day exit applies and the system keeps running on local hardware under the perpetual license described below.
The monthly fee covers a bounded scope:
What sits outside the monthly fee. Capital projects, major expansions, replacement of incumbent systems, and extensive specialist labor are quoted separately. This keeps the flat monthly fee predictable and large one-off work priced on its own merits.
Why a flat monthly, and where it goes. The effort is front-loaded. The first months carry the most engineering: standing up the pipeline, tuning it to the building's cameras, and wiring it into the desk. That effort tapers as the system stabilizes, while the value stays constant. The fee also carries recurring AI model and inference costs. For comparison, a single consumer AI subscription runs about $200 a month, on plans that train on your data. Elberta's stack instead runs on providers that do not retain or train on it, alongside hosted inference and the on-prem GPU. The flat monthly spreads the heavy early work across the term, so the board gets one predictable number and a system that keeps improving.
A simple way to view this long-term model is a fractional AI security operator for the building, one predictable number in place of variable license-plus-maintenance spend.
During the pilot the hardware is Elberta-owned. Elberta supplies it, carries the capital risk, and can recover or redeploy it if the pilot ends. The board's stake is in the outcome, and the equipment stays Elberta's through the pilot.
Camera count and expansion hardware. At the managed fee, Elberta will service up to about 12 cameras. The box Elberta supplies is sized for 12; any camera beyond 12 runs on hardware the association procures and owns, at cost. This still differs from the incumbent model: IntuVision charges for hardware and a recurring per-camera license both, while here the only added cost is association-owned hardware, with no per-camera license layered on top. This comparison holds where an open-source alternative performs satisfactorily on the cameras in question.
If the pilot proves out and the managed service continues, the box stays under the managed service as Elberta-owned equipment. The board's standing option to buy the hardware at Elberta's cost, described below, remains open at any time.
The association keeps its own footage and data; none of it is locked to Elberta. Elberta grants the association a perpetual, non-exclusive license to run the deployment built for the building for its own use, so the system keeps operating on local hardware even if the managed service ends. The board also has a standing option to buy out all hardware at Elberta's cost. What the monthly fee buys is ongoing tuning, maintenance, and improvement, not access to a locked portal. Elberta retains ownership of the reusable framework, integrations, and tuning logic it carries across deployments. Gallery is Elberta's founding reference implementation.
Elberta proposes to be the board's AI-powered managed security partner, starting with a $2,000 fixed 60-day pilot on Elberta-owned hardware, with the exit stated up front, and continuing on success at $1,000 per month for a defined scope. The next step is the board's decision to fund the pilot and/or a request for a demo of the proof of concept.