Easycoder Open Model Deployment and Operations on AWS

The problem this solves

Amazon Bedrock covers the popular models in the popular regions, and where it fits it is the right answer. Four cases fall outside it:

• An architecture Custom Model Import does not support. Gemma and its fine-tunes, including the medical variants, cannot be imported, so a team that has fine-tuned one has nowhere managed to put it.
• A region where the model is not offered. Custom Model Import is not available in Europe (London), so a UK organisation with a residency requirement and a custom model has no managed path.
• A model the catalogue does not carry at all: domain-specific and specialist open models, embedding models, speech models, and anything released in the last few weeks.
• Data that cannot leave your account: clinical records, financial files, legal documents, anything under a residency or contractual restriction.

Self-hosting usually means an engineering team learning GPU serving, quantisation, autoscaling and upgrade mechanics on the job, with an idle accelerator bill while they do it. Easycoder does that work as a fixed-scope engagement in your account, then operates the result.

How we deliver

• Assessment. One workload. Its job is to tell you whether you need this at all: if Bedrock already serves your model in your region we will say so and recommend Bedrock. Otherwise you get benchmarks of the open-weight candidates on the instance types that meet your latency target, measured cost per request at your volume, a residency and compliance position, and a written go or no-go recommendation.
• Deployment. One model into your AWS account to an agreed latency and cost target: model and instance selection, serving stack, autoscaling, region pinning, encryption with customer-managed keys, a per-request audit trail, load testing, runbooks and handover.
• Operations. Monthly. Health and latency monitoring, version and driver upgrades with rolling deploys, incident response, cost regression testing when a model or dependency changes underneath, capacity review, and a monthly report.

What you get

• The endpoint runs in your AWS account, under your IAM, your keys and your VPC. Your data does not traverse a shared endpoint or leave your account.
• Accelerator choice matched to the workload and to what it actually costs: Inferentia and Trainium where the model is supported, GPU where it is not.
• Acceptance is a measured benchmark against the target agreed up front, not an opinion.
• Region pinning, including Asia Pacific (Sydney), Asia Pacific (New Zealand), Asia Pacific (Melbourne) and Europe (London), including where the managed catalogue cannot host a custom model.
• Runbooks, infrastructure as code and a handover session, so your team can run it without us if you choose to.

Compliance frameworks

We design deployments against the framework your auditor actually uses, and we do not claim certifications we do not hold.

• New Zealand: Privacy Act 2020 and the Health Information Privacy Code 2020, HISO 10029 for health information security, and NZISM controls where government-facing work requires them.
• Australia: Privacy Act 1988 and the Australian Privacy Principles, the Australian Government Information Security Manual and the Essential Eight, APRA CPS 234 for financial services, and AWS Regions assessed under IRAP.
• United States health: AWS HIPAA-eligible services, operated under your own business associate agreement with AWS.
• Evidence you receive: a data flow and residency statement, a control mapping to the framework you name, input to your privacy impact assessment, and per-request audit and access logs.

Sector experience

Easycoder is based in Auckland. We have taken Generative AI workloads into production in Asia Pacific (New Zealand) and Asia Pacific (Sydney) for healthcare and financial services customers, with privacy impact assessments, independent penetration testing, clinician approval gates, customer-managed encryption and per-request audit trails. Typical starting points are a Gemma-family fine-tune no catalogue will host, a frontier API bill on one high-volume task, an embedding or speech workload at volume, or a clinical workload such as medical dictation and note structuring.

Support

Enquiries through this listing are answered within two business days, with a scoping call and a written proposal. After purchase: a named engagement lead, weekly status reporting, and support during New Zealand business hours. Response targets by severity are in the statement of work attached to each offer.

Documentation

Service descriptions and case studies: https://www.easycoder.co 

Infrastructure costs

The deployment runs in your AWS account. AWS charges for compute, storage and networking are billed by AWS to your account, separately from the price of this engagement, so you keep the capacity commitment and the accelerator discounts. Access is through IAM roles with least-privilege policies that we document and hand over.