BOBER AI
← All solutions
Self-hosted

Self-hosted AI

Full AI stack in your datacenter or private cloud: inference, RAG, UI — no foreign API dependency.

Message on Telegram

Self-hosted AI deploys full inference, embeddings and vector search on client infrastructure. Bober AI Systems designs for Yandex Cloud isolated, Selectel, bare metal in Moscow.

Components: vLLM/Ollama for LLM, sentence-transformers for embeddings, Qdrant/pgvector, Open WebUI, optional LangGraph agents. All in Docker/K8s with CI/CD.

Sizing depends on model (7B vs 70B), QPS and latency SLA. We do not oversell GPU — TCO calculated on Discovery.

Model updates via controlled rollout with eval gate. Air-gapped: offline model packages.

Links to enterprise processes: RAG on policies, CRM agents, document OCR — one private contour.

Infra + deploy budget from €5,000. Ongoing hosting and SLA optional.

When self-hosted is required

  • —Compliance blocks outbound calls to OpenAI/Anthropic
  • —Air-gapped or classified internal network
  • —Predictable TCO needed at high token volume
  • —SaaS AI vendor lock-in unacceptable

Deliverables

  • —GPU/CPU sizing and LLM deploy (vLLM, Ollama, TGI)
  • —Embedding service + vector database
  • —Open WebUI or API gateway for consumers
  • —Monitoring, backup, disaster recovery runbook

Phases

01

Requirements

Models, QPS, latency, network zones, compliance.

02

Hardware/cloud

GPU selection, K8s cluster, storage.

03

Stack deploy

LLM, embed, vector, gateway, UI.

04

Validation

Load test, failover drill, documentation.

Self-hosted stack

  • —K8s cluster with GPU node pool
  • —vLLM / Ollama inference services
  • —Embedding microservice + Qdrant
  • —API gateway + Open WebUI
  • —Prometheus/Grafana monitoring

Impact

100%

data residency control

−40%

token cost vs public API at scale

4–6 wks

typical full stack deploy

Related case studies

Email AI Assistant — Kaspersky

Email AI Assistant — Kaspersky

−50% повторных обращений L1

LLM · RAG · JavaScript

Консультанты быстрее закрывают типовые обращения, единообразно отвечают на вопросы о продуктах и тратят больше времени на сложные инциденты и апсейл.

View case study →

FAQ

Minimum hardware?
7B Q4 on 1× 24GB GPU for pilot. Production sizing individual.
Russian models?
Saiga, licensed GigaChat weights, multilingual embed.
HA?
Multi-replica inference, DB backups, K8s autoscaling.
MLOps team?
Runbook handover. SLA covers ops optionally.
Cloud vs bare metal?
Cloud faster start. Bare metal lower latency TCO at scale.
Security audit?
Network segmentation, pen test coordination on request.

Related services

Contact

Or leave a request

Full AI stack in your datacenter or private cloud: inference, RAG, UI — no foreign API dependency.