Boundary
Protected data never reaches a third-party AI API
Inference happens on agency-controlled GPUs. There is no external model endpoint in the path for protected records, and the inference VLAN has no default route out.
Public safety AI that stays inside the building.
Open-weight frontier models on hardware physically installed at the agency — from a single Mac Studio to NVIDIA Blackwell racks. Grounded in the sources you approve. Nothing hosted by anyone else, and no protected data sent to third-party AI services.
0third-party AI APIs in the path for protected data
4tiers chosen by use, priced in the open from about $25,000 USD
100%of weights, indexes, and logs stay on agency premises
Everyanswer cites its sources and is logged for review
Why on site
Cloud AI asks agencies to send data out. This does not.
Police services hold occurrence data, personnel records, and community contacts that should not leave their control. Most AI products require exactly that. Private AI runs the models where the data already lives — on an inference network with no path to the public internet.
Control
The agency decides what the AI can see
Models are grounded only in sources the agency approves, with access scoped by role like every other platform service.
Continuity
Works without an outside dependency
No usage-based API pricing, no vendor model changes overnight, no outage in someone else's region. Model files are pinned by hash on the purchase order.
What a member actually sees
The passage. The source. Then a person decides.
Ask a question inside 10-8 or HRIS and get the relevant section from your own procedures, with the source attached. Nothing leaves the building to produce it, and every query is logged.
- Grounded in policies, procedures, HRIS records, and 10-8 data the agency approved
- Cited so a supervisor can open the exact document behind the answer
- Scoped to what the member could already open directly
One Bridge 10-8 · Private AIOn-site
Sgt. MoreauWhich procedure applies when a paid-duty client is a licensed event serving alcohol?
Licensed events fall under Paid Duty Procedure 4.12 — Licensed Premises. A minimum of two officers is required when expected attendance exceeds 250, and the client must confirm licence details on the request before approval. The supervisor of record signs off on staffing before the opportunity is posted to members.
Logged · model pinned · role: SupervisorReviewed by a person before action
Four tiers, chosen by what you want the AI to do
Start with your 1 Bridge Tec products. Connect the rest when you are ready.
Every tier is hardware One Bridge Tec installs inside your building. One Bridge Tec does not host your AI, and neither does anyone else: your premises, your data, your keys, your audit log. The tiers differ only in what the AI is connected to and how much of the agency it serves.
Same at every tier
- No protected data sent to any third-party AI service
- Agency-owned or leased hardware on agency premises — never a hosted or shared facility
- Role-scoped access, cited answers, and a full query log the agency keeps
- Open-weight models pinned by hash — the agency decides when anything changes
Tier 1 · Entry
Studio Assist
One Mac Studio on site for a smaller service — about 250 officers or fewer — running 1 Bridge Tec products only.
$25,000 – $32,000
USD, installed, est. 2026. Apple has not yet published the 512 GB price; band assumes above the $18,299 list for 256 GB / 16 TB.
Use it for
- Score and rank résumés against the posting, with reasons a panel can check
- Enhance job descriptions and postings in the agency's own voice and policy
- Help build schedules and fill callbacks, paid duties, and overtime
- Ascertain coverage for an event: who is eligible, who is available, where the gaps are
- Answer procedure and policy questions with the source passage attached
Hardware installed
- 1× Mac Studio, M5 Ultra: 36-core CPU, 80-core GPU, 32-core Neural Engine
- 512 GB unified memory at 1.2 TB/s, 16 TB SSD — largest configuration Apple offers
- 10 GbE, Thunderbolt 5, 1.5 kVA UPS, locking cabinet
- Optional second Mac Studio clustered over Thunderbolt 5 for capacity or standby
- Power
- Under 300 W. Any standard outlet in a locked room.
- Runs
- 70B dense and 120B mixture-of-experts models for a handful of people at once, plus queued batch work such as scoring an applicant pool.
Tier 2
Platform Assist
The same assistance as Tier 1 on NVIDIA hardware, for a larger service or more people working at once — HRIS and 10-8 only.
$35,000 – $70,000
USD, installed hardware, est. 2026
Use it for
- Score and rank résumés against the posting, with reasons a panel can check
- Enhance job descriptions and postings in the agency's own voice and policy
- Help build schedules and fill callbacks, paid duties, and overtime
- Ascertain coverage for an event: who is eligible, who is available, where the gaps are
- Answer procedure and policy questions with the source passage attached
Hardware installed
- 1–2× NVIDIA RTX PRO 6000 Blackwell GPUs, 96 GB each
- Certified workstation-class server, 512 GB system memory, NVMe storage
- 3 kVA UPS, 10 GbE switch, locking cabinet
- Runs from a standard 20 A circuit
- Power
- 1.0–1.8 kW. Normal 120–208 V outlet in a locked room.
- Runs
- 70B dense and 120B mixture-of-experts models for a service using HRIS and 10-8.
Tier 3
Agency Connect
Everything in Tier 2, plus the systems you already run in house, connected on the same private network.
$220,000 – $340,000
USD, installed hardware, est. 2026
Use it for
- Connect records management, CAD exports, document repositories, file shares, and training records
- Search and summarize across systems with every answer cited back to its source
- Draft from the agency's own precedents: reports, briefings, correspondence, disclosure prep
- Serve 30–80 people at once across headquarters and VPN-connected detachments
- Redact personal information before a prompt ever reaches the GPU where policy requires
Hardware installed
- 8× NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, 768 GB GPU memory
- Dual-socket 4U server, 1–2 TB system memory, 60 TB NVMe plus 100 TB storage for weights and audit logs
- 100 GbE networking, 42U rack, 15–20 kVA online UPS
- Cooling upgrade where the comms room needs it
- Power
- 5–7 kW. Requires 208 V / 30 A+ service, not a wall outlet.
- Runs
- Many concurrent 70B and 120B replicas and long-context models across connected systems.
Tier 4
Sovereign Scale
Everything in Tier 3 for a regional, state, or provincial service, several agencies, or the largest open-weight models — in a data hall the agency owns.
$480,000 – $650,000
USD per production node, installed. HA pair $950,000 – $1,300,000
Use it for
- Multi-agency tenancy with hard separation between each service's data and users
- The largest open-weight models for long documents, complex analysis, and 100+ users
- Private fine-tuning on de-identified agency material, on the agency's own GPUs
- High availability with a second identical node — no dependence on any outside provider
- Requires a data hall the agency owns: a headquarters facility, a state or provincial police data centre, or a room One Bridge Tec helps you build out
Hardware installed
- 8-GPU NVIDIA HGX or DGX B200, about 1.4 TB of HBM3e connected by NVLink
- InfiniBand or 400 GbE fabric, 200 TB+ object store for weights and archives
- Second identical node recommended for high availability
- Installed in an agency-owned data hall with 15 kW per rack, cooling, and generator-backed power
- Power
- 12–15 kW per chassis. Needs an agency-owned data hall, not a station comms room.
- Runs
- 671B-class mixture-of-experts models with usable context, 400B+ models, and many 70–120B replicas at API-like latency.
Figures are 2026 list and street bands in US dollars for installed hardware, before tax, installation labour, and training. They are planning numbers, not a vendor quote; GPU allocation and exchange rates move. Local taxes, exchange rates, and installation labour are added at scoping. Three-year all-in estimates including power and support: Studio Assist about $45,000; Platform Assist about $80,000; Agency Connect about $420,000; Sovereign Scale about $900,000 per node, excluding the agency's own facility costs. Tiers can be stepped up in place: the boundary, models, and logs carry forward.
The alternative
For a smaller service, the other option is an API to Anthropic, OpenAI, or xAI.
Those are capable models. The difference is where your information goes and who sets the rules.
Third-party AI API
- Every prompt, résumé, schedule, and policy excerpt is transmitted to the vendor's servers, outside the agency's control and often in another jurisdiction
- Retention, staff access, and acceptable use are governed by the vendor's terms, which can change
- Billed per token; costs rise with use and are hard to budget
- Models are updated or retired on the vendor's schedule
- An outage or policy change on their side stops your work
Private AI on site
- Prompts and records never leave the machine in your building
- Retention, access, and audit are set by the agency in its own policy
- One-time hardware cost or a lease; no per-token bill
- Model files pinned by hash; the agency decides when anything changes
- Works with no internet path at all
Not sure which tier?
A 30-minute scoping call settles it.
We look at what you want the AI to do first — hiring, scheduling, event coverage, or connecting systems you already run — plus your users and what your building can power. You leave with a tier, a price band, and a clear next step. No procurement commitment.
Book a scoping call
Email [email protected]
Prefer a demo first? Book a demo and we will show the cited-answer experience on a live cluster.
The platform
Enterprise inference on NVIDIA Blackwell
Each deployment is configured to order and installed by One Bridge Tec. Hardware is ordered per customer from certified OEM partners, staged, and commissioned on an isolated inference network before the first query.
Agencies can purchase the hardware outright or lease it so the cost model matches capital or operating budgets. One Bridge Tec does not host the hardware and does not offer a data-centre option: every tier is installed at premises the agency owns and controls.
- Compute
- Mac Studio M5 Ultra with 512 GB unified memory for Studio Assist; NVIDIA RTX PRO 6000 Blackwell for Platform Assist and Agency Connect; HGX or DGX B200 with NVLink for Sovereign Scale.
- Serving
- MLX on the Mac Studio; vLLM or TensorRT-LLM with continuous batching on NVIDIA. Both sit behind the same self-hosted, OpenAI-compatible gateway that routes only to local hardware.
- Network
- Separate inference, management, and storage VLANs. No default route to the internet on the inference VLAN. Hardware firewall; optional one-way log export.
- Storage
- Local NVMe for hot weights, a local model registry and audit archive, and offline encrypted copies of every production weight hash.
- Identity
- Keycloak or the agency's own on-premises directory. Every request carries the user, time, model hash, and retention policy.
- Procurement
- Purchase or lease. Ordered per customer and installed on the agency's premises by One Bridge Tec. Never hosted.
Who it is for
Built for the people who carry the responsibility.
Private AI is not a public chatbot. It assists authorized agency staff inside the One Bridge Tec products they already use.
Chiefs and executives
Answer the board's first question before it is asked.
Where does the data go? Nowhere. Private AI gives leadership a defensible answer on residency, audit, and cost — with a published price band instead of a vendor mystery.
How leaders use One Bridge Tec
Records, HR, and analysts
Draft in minutes from the documents you already own.
Job postings, panel summaries, policy answers, and disclosure prep drafted from agency sources with citations you can check — never from a public model that saw your files.
See 1 Bridge Tec HRIS
Officers and supervisors
Ask how a procedure applies and get the passage, not a paraphrase.
On a phone or in 10-8, members get the relevant section with its source. Supervisors see readiness conflicts flagged before a paid duty or callback is offered.
Explore One Bridge 10-8
The models
Open-weight frontier models, adapted for public safety, grounded in your data
Closed models from OpenAI, Anthropic, and Google cannot be run privately; those vendors do not ship weights. Private AI runs the latest open-weight frontier models locally. One Bridge Tec adapts them for public safety language, procedure, and document types, then grounds every answer in the agency's approved sources so responses cite what they draw on.
Foundation
Available
Open-weight frontier models run locally
DeepSeek, Qwen, GLM, Kimi, and MiniMax-class models are downloaded once, verified by hash, and run entirely on the agency's GPUs.
Adaptation
Available
Public safety adaptation
Prompting, evaluation sets, and retrieval tuned for policing procedure, HR, and community-service contexts.
Grounding
Available
Retrieval over agency-approved sources
Policies, procedures, HRIS records, and 10-8 operational data — each collection permissioned by role.
Citations
Available
Answers show their sources
Every response links back to the documents or records it used so a person can verify it.
Tuning
In development
Agency-specific fine-tuning
Optional private fine-tuning on de-identified agency material, performed on the agency's own GPUs, only with a written legal basis.
Evaluation
In development
Continuous quality checks
Agency-defined test sets run against every model or prompt update before it is promoted.
What agencies use it for
Assistance across the One Bridge Tec platform
Operations
10-8 staffing assistance
Summarize open paid duties, flag readiness conflicts, and draft client updates from the operational record.
Learn morePeople
HRIS drafting and search
Draft job postings, summarize candidate files for panels, and answer policy questions from the agency's own documents.
Learn morePolicy
Procedure and policy lookup
Members ask how a procedure applies and receive the relevant passage with its source, not a paraphrase.
Community
Anonymous service triage support
Assist staff handling community contacts while the anonymity boundary of the community platform is preserved.
Safeguards
Private is necessary. It is not sufficient.
Keeping data on site removes the largest exposure. The rest of the platform's controls still apply — and people stay in charge of every decision.
Assistance, not autonomous decisions
The system drafts, summarizes, and retrieves. People decide. It does not make staffing, hiring, or enforcement decisions.
Role-scoped retrieval
A member cannot retrieve through the AI what they could not open directly. Personal information is redacted before a prompt reaches the GPU where policy requires.
Every query and source logged
User, time, model hash, prompt, retrieved documents, and response are recorded, with retention set by the agency.
Encryption, pinned weights, physical control
Encrypted drives, change control on every weight file, signed updates applied by One Bridge Tec, and hardware behind badge access and cameras.
Deployment
From order to first query
01
Scope
Confirm what the AI should do first, which systems it may see, who uses it, and what the building can power. Choose a tier and purchase or lease. Everything is installed on your premises.
02
Order and stage
Hardware is quoted through certified OEM partners, ordered for the agency, and configured by One Bridge Tec before delivery.
03
Install and connect
Installed with its isolated inference network, joined to the agency directory, and connected to approved sources.
04
Evaluate and go live
Agency test sets run; roles, audit, and the pinned model hash are confirmed; assistance is enabled by module.
Common questions
Private AI questions
Does any data ever leave the agency?
Protected agency data is processed on agency-controlled GPUs and is not sent to third-party AI services. The inference network has no path to the public internet; updates arrive through a locked management host or offline media. Weights, indexes, and logs stay on the agency's premises.
Which tier do we need?
A service of about 250 officers or fewer that runs 1 Bridge Tec HRIS or 10-8 and nothing else can start with Studio Assist: one Mac Studio on site for résumé scoring, job descriptions, scheduling help, event coverage, and policy lookup. Larger services, or more people working at once, move to Platform Assist on NVIDIA hardware for the same assistance. Choose Agency Connect when you want the AI to draw on systems you already run in house — records, CAD exports, document repositories. Sovereign Scale is for regional or multi-agency services, the largest models, and full high availability. Every tier has the same privacy boundary; you can step up in place.
When is a Mac Studio enough, and when do we need NVIDIA?
A Mac Studio with 512 GB of unified memory holds 70B and 120B-class open-weight models entirely on one desk-sized machine and serves a handful of people at once — enough for a smaller service using 1 Bridge Tec products, with batch jobs such as scoring an applicant pool queued rather than instant. Once dozens of people need answers at the same time, or you connect records, CAD, and document systems, you need NVIDIA GPU memory, bandwidth, and continuous batching. The software, boundary, and logs are the same; only the hardware changes.
Can we lease instead of buy?
Yes. Leasing is available so the hardware can be treated as an operating expense. Leased or purchased, the hardware is installed on the agency's premises. One Bridge Tec does not host it.
Do you offer hosting in a data centre?
No. The point of Private AI is that the hardware sits where the data already lives, under the agency's physical control. Sovereign Scale needs a proper data hall because of its power and cooling, but that hall belongs to the agency — a headquarters facility or a state or provincial police data centre — not to One Bridge Tec or a colocation provider.
Can it run GPT, Claude, or Grok on site?
No, and no vendor can offer that. OpenAI, Anthropic, xAI, and Google do not ship their model weights; the only way to use them is their API, which means sending your prompts and records to their servers. Private AI runs the leading open-weight models, which are the only frontier-class models that can be operated privately.
Is this a chatbot for the public?
No. Private AI assists authorized agency staff inside One Bridge Tec products. It is not exposed to the public.
Are the prices a quote?
No. They are 2026 planning bands for installed hardware in US dollars, before tax, installation labour, and training. Apple has not yet published the 512 GB Mac Studio price; GPU allocation and exchange rates change. One Bridge Tec confirms pricing at scoping.
Related
Bring enterprise AI inside the agency.
Scope a private AI deployment with the team that builds the platform it runs on.