--- title: "Private AI deployment inside your boundary | YPAI" url: https://ypai.ai/ai-implementation/private-enterprise-deployment/ description: "Assistants, agents, document workflows and private models, built by YPAI and running inside your own cloud, a dedicated EEA tenant or your own hardware." source: "src/copy/routes (route /ai-implementation/private-enterprise-deployment/)" --- # The pilot ran on a laptop. Production runs in your building. > Assistants, agents, document workflows and private models, built by YPAI and running inside your own cloud, a dedicated EEA tenant or your own hardware. Assistants, agents, document workflows, voice and private models. Evaluated on your data before release and run by YPAI. The data, the region and the deletion stay yours. Try the assistant in your browser - **Your cloud**: AWS, Azure or Google, your account - **EEA tenant**: dedicated, run by YPAI - **Hybrid**: records inside, context out - **Your hardware**: disconnected once tested Built by YPAI. Runs in your browser The result shown is the bundled example. Run it to see your own. No customer data. Live from the EU inference profile in Frankfurt {n} held-out items · {critical} critical one frame · snowfall · dusk ## Your case files stay in the archive. The assistant reads them there. Three documents, one question. The assistant finds the passage that answers it and quotes it back, so the reviewer can check the source. Paste your own documents and ask your own question. - Change control procedure, section 4. Emergency changes to the inference service may be applied by the on-call engineer after approval from the duty manager, and must be recorded in the change log within 24 hours. - Retention schedule, row 7. Prompts and responses on the managed model path are retained for zero days. Application logs are retained for 400 days in the customer's account and deleted by the customer. - Access policy, section 2. Support access to the production environment is granted per incident by the customer's platform owner and expires after eight hours or at incident close, whichever comes first. The demo models are trained on English. On documents in other languages the answer can be imprecise. Run in your browser (40 MB download) This counter starts in your browser when you run the demo. - [Knowledge assistants](https://ypai.ai/enterprise-knowledge-assistants/) ## Four places it can run. One boundary you sign. You choose the pattern. The boundary is drawn and counted around it, and your security owner signs it before release. The same systems run in all four. Swipe sideways for the other patterns. - Your AWS, Azure or Google account, your region, your identity provider. Model calls leave only over private endpoints you approve, with zero retention on the far side. - A dedicated tenant inside EEA jurisdiction, run by YPAI. Keys are generated and held by you. Region and deletion stay your decisions. - Records, retrieval, permissions and actions stay inside. A frontier model outside answers only from context the gate approved. - Weights, runtime, application and index on hardware you own. YPAI calls a deployment disconnected only after inference, telemetry, updates and support access are shown to stay inside. We read the pattern from your brief and check it with you in our reply. Every pattern ships with the same documentation and the same service levels. ## Two paths leave the building. Both are named. This is one deployment's map. Two private paths leave it, one to the model and one to monitoring, and the count of unapproved paths reads zero. Switch a public path on and watch the count move. Swipe sideways to see both sides of the map. Yours is drawn at scoping and signed before release. ## Every record has a person in it. The workflow holds the person and moves the record. An admission note goes through the workflow. The names, places and dates are marked and kept inside. The clinical facts and the task may cross to a model outside. Patient Ola Nordmann, born 12 March 1961, was admitted to Ullevål on 4 May with chest pain. Dr Anne Berg ordered a troponin test and an ECG. Summarise the admission for the discharge letter. Run in your browser (13 MB download) Extracted for the system of record - **patient**: held - **admitted**: Ullevål, 4 May - **presenting**: chest pain - **ordered**: troponin test, ECG - **task**: discharge letter summary Run it and the names, places and dates are marked on the record and kept inside. The rest may cross. 5 held inside: 2 people, 2 dates, 1 place. The rest may cross to a model outside. the bundled example, until you run your own {n} held inside: {counts}. The rest may cross to a model outside. Nothing held. The whole record may cross. The demo model that finds names and places is trained on English text. Dates are found by a rule. Illustrative record. No patient or case is real. The record stays where it is stored, and the workflow reads it there. - [Document workflows](https://ypai.ai/document-ai-workflow-automation/) ## The demo answered the demo questions. Release waits for your cases. Before release the system is evaluated inside your boundary on sets built from your production data, with hard variants of your own questions. Here the gate runs on eight held-out items. Two are critical. Every critical item passes and the pass rate is at least 75 percent. Run in your browser (14 MB download) the bundled example set, until you run the gate The demo reader reads English only, so the gate runs each item in its English version. ### Who approves an emergency change? ### Within what time must the change be recorded? ### Who applies the change? ### How long are prompts retained on the managed model path? ### Who deletes the application logs? ### For how long, if at all, does the managed path keep what was asked? ### When does support access expire if the incident runs for a day? ### Which role can grant production support access? ## Your users speak in corridors and cabins. The transcript stays with you. Record three seconds in the room you are in. The take is transcribed in your browser. In your deployment the transcript goes straight to your workflow. start route guidance to the harbour terminal The demo model transcribes English speech only. - [Speech data](https://ypai.ai/speech-data/) ## The camera sees the crossing at dusk. The reading is made on your hardware. One frame from a street camera in snowfall. The detector marks what it finds, with its confidence. In your deployment the frame stays on your hardware and the operator accepts or corrects the detection. Run in your browser (3 MB download) the detector marked what it found person 0.68. The operator accepts or corrects it before the detection is recorded. the bundled example, until you run the detector on this frame {readings}. The operator confirms or corrects it before the reading is written. The detector found nothing above the threshold in this frame. - [Industry solutions](https://ypai.ai/industry-solutions/) ## Your users write in Norwegian. The model answers in it, from Frankfurt. Open-weight and fine-tuned models in the Llama and Mistral families, evaluated for the languages your users speak, served from Frankfurt on the EU inference profile, so prompts are processed in the EU. The agent acts only inside the permissions granted for the task and shows every step. ### A private model in Norwegian Oppsummer denne saken for saksbehandleren i to setninger. Søknaden gjelder tilknytning til kommunalt vann og avløp for eiendommen i Fjellveien 12, og mangler situasjonskart og samtykke fra nabo. Saken settes på vent til begge dokumenter er mottatt. A worked example, not a live run. - [AI Implementation](https://ypai.ai/ai-implementation/) ### An agent inside its permissions Check the incident against the runbook and file the change request. - change request drafted, rollback path named - sent to the duty manager, waiting Filing the change is outside the permissions granted, so the run stops there and the trace shows where. - [Enterprise automation](https://ypai.ai/enterprise-automation-solutions/) ## The review asks three questions. The deployment answers them. Most private deployments stop at the same three questions. Each has a document behind it, produced as part of delivery. The review that stopped the pilot is the one the pack is written for. Not yet run on this page Switch a path on the map above ↑ Handed over in the documentation below ↓ - delivery, monitoring and managed improvement under an agreed scope - every release through the gate, critical items first - rollback to the last known good version, five-minute target - the runbook and the escalation paths, defined before production - the data, the region and the deletion - approval limits for every agent, actions fail closed outside them - **Egress map**: What leaves our environment?. Every path named, approved or closed. Zero unapproved paths at release, counted, with the retention rule per component. - **Acceptance report**: Does it work on our data?. Evaluated inside your boundary on your production data, with hard variants and a gate that blocks on critical criteria. - **Runbook and ownership**: Who runs it after go-live?. YPAI runs delivery, monitoring and managed improvement under an agreed scope. Configuration authority, data, region and deletion stay yours. ## Twelve numbers. The SOW binds each one. Service specifications for a named deployment profile. The statement of work names the target, the architecture, the prerequisites and the measurement method. Production starts after the acceptance evidence passes. - **Network boundary**: unapproved egress paths. data-boundary. 0 - **Inference retention**: days of prompt and response retention. data-boundary. 0 - **Availability**: monthly service availability. release-and-operations. 99.99% - **Recovery**: critical writes, RTO under 60 s on the serving path. release-and-operations. RPO 0 - **Private 70B inference**: p95 to first token, 50 ms per token. models. 1 s - **Fast 8B inference**: p95 to first token, 10 ms per token. models. 250 ms - **Release quality**: required gates, 0 critical findings. release-and-operations. 100% - **Rollback**: to the last known good version. release-and-operations. 5 min - **Audit and keys**: agreed event coverage, customer-managed keys. data-boundary. 100% - **Environments**: isolated environments. release-and-operations. 3 - **Actions**: outside approved authority. data-boundary. fail closed - **Model control**: targets only. models. policy-equivalent The statement of work binds the selected profile. Production begins after its prerequisites and acceptance tests pass. ## One pack per workload. Handed over at release. Written for the people who sign. {Documents} documents, in the order your review asks its questions, and {numbers} of them carry a service level from the statement of work. one workload, the profile named on every sheet Signed by your security owner before release. service levels in the statement of work #### Data-flow map - Every path in and out, named, approved or closed, with the network zones. #### Capability matrix - Model, endpoint, tool, cache and logging, for every zero-retention path. #### Key and region matrix - Which keys are customer-managed, and where each component processes. #### Identity and agent scoping - Who and what may act, with the approval limit for every agent. #### Evaluation suite and acceptance report - Critical criteria and protected slices, run on your cases and reported before release. #### Provenance, SBOM and release evidence - Signed provenance, the bill of materials and the release evidence for every production artefact. #### Release and rollback procedure - How a release goes out, and how it comes back. #### Runbook, ownership and handover - Who runs what after go-live, and how it is handed over. ### What do we show the regulator? #### AI Act deployer evidence - Assigned oversight, monitoring and log retention, as the Act asks of a deployer. #### DPA and subprocessor list - The data processing agreement and every subprocessor named. The pack is the deployment on paper. The review reads it before production. ## Three lines are enough. What it does, where it must run, what must not leave. ### What you bring - The system, or the pilot and where it stopped. ### Where it must run - Your cloud, an EEA tenant, a hybrid split, or your hardware. ### What must not leave - The records, the index, the keys. A named project lead reads it and replies inside one EU business day with the pattern, the prerequisites and a first read on the boundary. What you bring, where it must run, what must not leave This is what we read in your brief. Correct it if it is wrong. A named project lead reads what you bring and replies inside one EU business day with the pattern, the prerequisites and a first read on the boundary. The boundary you sign starts as these three lines. ## What security, compliance and platform ask first. ### Can the complete system run inside our environment? Where the selected models, licences, infrastructure and dependencies support it. Inference, retrieval, application state, logging, updates and operational access are checked against the same service levels: 0 unapproved egress paths and an explicit retention rule per component. A system is described as fully private only when every material process runs inside. ### Can sensitive data remain inside while an external model is used? In a hybrid design, source documents, retrieval, permissions and actions stay inside and only approved context reaches the external model. That hop uses 0-day prompt and response retention while stateful components keep their own approved storage and deletion rules. ### Do you support on-premises or air-gapped environments? Where the model, licence, hardware and update process permit it. YPAI calls a deployment disconnected only after inference, telemetry, package retrieval and support access are shown to stay inside the boundary, with unapproved egress at 0. ### Our pilot stalled at security review. Where do we start? With the boundary. We draw it for your pattern, count what crosses, and hand security the egress map and the retention and key matrix. The pilot is then evaluated inside it on your own cases before anyone decides on production. ### Who operates the system after launch? YPAI can transfer the deployment to an internal owner or continue under an agreed monitoring and update scope. Ownership, access, the availability profile and its prerequisites, RTO, RPO and escalation paths are defined before production. ### Does private deployment make the system GDPR compliant? Location, 0 unapproved egress paths and a retention matrix support the applicable obligations. Roles, purposes, transfers and subprocessors belong in the DPA and the organisational framework, which the deployment documentation carries. You name what you bring and where it must run. We build it there, test it there, and run it with you. AI Implementation covers discovery, architecture and the deployment itself.