openjevx Server - self-hosted decision-model server

One-click launch (us-east-1; pick your Region in the console): https://console.aws.amazon.com/cloudformation/home?region=us-east-1#/stacks/quickcreate?templateURL=https://deemwar-marketplace-media-663532473958.s3.amazonaws.com/jev/cfn/openjevx-server-ecs-0.5.11.cfn.yaml&stackName=openjevx-server 

BEFORE YOU LAUNCH You need an OpenJevX-format model in S3 in the same account: a model.tar(.gz) holding model/ (openjevx.w8.onnx, config.json, tokenizer.json), or a prefix holding those three files. openjevx Server serves OpenJevX-format models only, not arbitrary ONNX models.

LAUNCH (about 10 minutes)

1. Choose "Continue to Launch" > delivery option "openjevx Server on Amazon ECS Fargate" > Launch CloudFormation.
2. Parameters:
   • ModelUri: s3://<bucket>/<path>/model.tar (or s3://<bucket>/<prefix>/). The task gets read access to that bucket only.
   • ApiKey: 24+ characters. Clients send it as "Authorization: Bearer <ApiKey>" on every /v1 call.
   • DashboardPassword: 12+ characters (no double quote or backslash).
   • Scheme: "internal" (recommended; reachable only inside the new VPC) or "internet-facing".
   • AllowedCidr: who may reach the load balancer (default: the VPC only).
   • TaskCpu / TaskMemory: 2048 / 4096 recommended.
3. Create the stack. When it is CREATE_COMPLETE, the Outputs tab shows Endpoint. The model loads 1-2 minutes after the task starts; /health then shows its version and sha256. Without a readable model the server does not start.

USE curl -s <Endpoint>/health curl -s <Endpoint>/v1/systemone -H "Authorization: Bearer <ApiKey>"
-d '{"state": {"order_total_rs": 501}, "questions": {"discount": {"type": "noul", "instructions": "Does the order get the discount? Orders over Rs 500 get 10% off."}}}' Every answer returns a probability per option. Types: noul (yes/no), choice (pick one), score (1..N levels). Calls without the right key get 401. Responses carry a Server-Timing header (server-side milliseconds).

UPDATING THE MODEL Overwrite the object at ModelUri. The server checks it every minute and swaps to the new model between requests - no restart, no dropped requests - keeping the previous model in memory. A bad upload is rejected and the current model keeps serving. Roll back by uploading the previous version again (S3 versioning recommended).

SUPPORT openjevx Server is free and open source (Apache-2.0). Support: io@deemwar.com .

SIZING AND COST

• CPU only. A 598 MB 8-bit model uses about 1.6 GB RAM.
• Server-side, in vendor tests on Fargate, per vCPU: 0.3-1.05 s for a short decision (50 tokens) and 3.6-10.5 s for a 512-token input; the range follows the host CPU Fargate assigns (one inference thread per vCPU is set for you). Add tasks (DesiredCount) for throughput.
• You pay only for your AWS resources (Fargate task, load balancer, logs).

REMOVE Delete the CloudFormation stack. Your model bucket is never touched.

SECURITY

• Runs as non-root on a distroless image.
• No outbound calls except reading your model from S3. No telemetry. Request bodies are never logged.
• Secrets live in AWS Secrets Manager. Support: io@deemwar.com