Skip to content
zinto

On-prem AI for data conscious industries

We build AI processes on private servers.

Production AI systems deployed on your own servers, on premises, for organisations that cannot send their data to public AI services.

Talk to our teamExplore services

Our team has worked at organisations such as

Your most valuable data is the data you cannot send anywhere.

OAIC guidance on AI productsThe OAIC recommends that organisations "do not enter personal information, and particularly sensitive information, into publicly available generative AI tools". So the AI comes to the data, on servers you own.Read the guidance

Seven services run on one stack

Each one installs like a unit in a rack. Hover or tap a unit, then open it.

U1Strategy and readiness auditU2On-prem AI deploymentU3Forward deployed engineersU4Agents and workflow automationU5Data engineeringU6LLM integration and RAGU7Model fine-tuning

U1 / Service

AI strategy and readiness audit

A vendor neutral audit of your data, systems and use cases, and a ranked roadmap of what to build and what to skip.

stack: Claude (Anthropic API), OpenAI, Llama, Mistral, AWS

Open the service →
Example build / document search for a regulated teamIllustrative

One GPU server in your server room. Sized from your real load before you buy anything.

2×NVIDIA L40S GPUs in a 2U server
96GBGPU memory: 2 × 48 GB
35GBWeights for a 70B model at 4-bit: 70B × 0.5 bytes
0Requests sent to a public AI service
Full arithmetic: how much GPU memory an LLM needsGPU power at full load: 2 × 350 W

Five steps. Four ways to engage.

Every engagement runs the same five steps. The engagement model decides which of them we own.

Forward deployed teamProject deliveryAdvisory retainerBuild, operate, transfer
  1. 01

    Discovery

    An intro call on your objectives and goals, then we sit with your team to learn the business and how the work runs today. We agree the decision the system supports and the one metric that says it works.

  2. 02

    Ground the data

    We find where the data lives, who owns it and whether it reconciles, and plan where the data and the model will sit.

  3. 03

    Prototype and evaluate

    We build on your real data and score the output against an evaluation set from your own past cases before it carries load.

  4. 04

    Integrate

    It goes into the workflow and your systems of record, with review points where an error costs money, monitoring and an audit trail.

  5. 05

    Operate and hand over

    We run it with your team, then hand over the code, the evaluation suite, runbooks and training. It is not finished until you are satisfied.

How we run each step
Case study / Sydney construction business
10,000 job PDFs nobody could search, made askable on a private server. Ten staff find in seconds what took hours of digging.
HostingA rented private server
System of recordProcore, unchanged
EstimatesDrafted in minutes, not hours
Read the case study →
Who we work with

FinanceBankingInsuranceReal estateAccountingAuditAsset and wealth managementPayments and fintech

ConstructionTradesLogisticsManufacturingWholesale

Hover an industry

The stack, as we ship it

zinto-stack.yaml
# runs on your servers. cloud regions are the fallback.
infrastructure:  [on-prem NVIDIA GPUs, bare metal, Kubernetes, Docker, Terraform]
models:          [Llama, Mistral, Claude, OpenAI, Hugging Face]
serving:         [vLLM, LangGraph, LlamaIndex]
data:            [PostgreSQL + pgvector, Qdrant, Elasticsearch, Kafka, dbt, Airflow]
quality:         [pytest, evaluation harnesses, OpenTelemetry, Langfuse, Grafana, MLflow]
standards:       [Privacy Act 1988, APRA CPS 234, CPS 230, AML/CTF Act, ISO 27001, Essential Eight]
egress:          deny

Tell us where your hours go.

We map the work, size the hardware and price what is worth automating.

Talk to our team →