AI Platform Infrastructure and Operations
Production-v2 runtime
Production-v2 runs in AWS account 564141170168, region ap-southeast-1.
| Component | Active production-v2 runtime |
|---|---|
| Controller | ECS cluster dmind_ai-service_ai-controller_ecs-cluster_prod-v2; service dmind_ai-controller_service_prod; image dmind-ai-service-ai-controller:v1.2.0; log group /prod-v2/ai-controller. |
| Converter | ECS service dmind_ai-service_converter_ecs-service_prod-v2; image dmind-ai-service-converter:prod_v1.1.2; internal ALB dmind-converter-alb-prod-v2; log group /prod-v2/converter. |
| DMIND ASR on-cloud | ECS service dmind_ai-service_asr-oncloud_ecs-service_prod-v2; image dmind-ai-service-asr:oncloud-dccd177; log group /prod-v2/asr. |
| DMIND ASR on-prem caller | ECS service dmind_ai-service_asr-onprem-caller_ecs-service_prod-v2; image dmind-ai-service-asr:onprem-969bd0d; log group /prod-v2/asr. |
| OpenAI ASR | ECS service dmind_ai-service-openai-asr_service_prod-v2; image dmind-ai-service-openai-asr:prod-v2; log group /prod-v2/ai-service_openai-asr. |
| DSC model API | ECS service dmind_ai-service_dsc_ecs-service_prod-v2; image dmind-ai-service-dsc:prod_1.0.0; log group /prod/dsc-api. |
| DSC invoker | Lambda dmind_ai-service_dsc_invoker-function_prod-v2, enabled through alias distributer -> version 7. |
| GPT worker | ECS service dmind_ai-service-openai-gpt_service_prod-v2; image dmind-ai-service-openai-gpt:v1.0.3; log group /prod-v2/ai-service_openai-gpt. |
| MongoDB | Atlas project DMIND Production; cluster dmind-ai-service-prod; database ai-service; collections prediction and openai-gpt-response. |
| Media | S3 dmind-aiservice-asrfilestorage-prod-v2. |
Queue topology
| Purpose | Production-v2 queue |
|---|---|
| ASR requests | dmind_ai-service_asr_queue_prod-v2.fifo |
| ASR tier-1 failures | dmind_ai-service_asr_dead-letter-queue-tier1_prod-v2.fifo |
| ASR tier-2 failures | dmind_ai-service_asr_dead-letter-queue-tier2_prod-v2.fifo |
| DSC requests | dmind_ai-service_dsc_queue_prod-v2.fifo |
| GPT requests | dmind_ai-service_openai_queue_prod-v2.fifo |
| GPT failures | dmind_ai-service_openai_dead-letter_queue_prod-v2.fifo |
| Unified worker results | dmind_ai-controller_result-queue_prod-v2.fifo |
Each environment uses a matching queue family, S3 bucket, MongoDB database, and worker configuration. Queue messages contain S3 paths and prediction IDs scoped to that environment. The on-cloud, on-prem caller, and OpenAI ASR services are competing consumers of the same ASR request FIFO; the controller does not select a worker implementation for each message.
Source-to-runtime ownership
| Runtime | Source repository or artifact status |
|---|---|
| AI controller | dmind-aiservice-aicontroller, release v1.2.0. |
| DSC API and invoker | dmind-aiservice-dsc. |
| GPT worker | dmind-aiservice-openai-gpt-golang, release v1.0.3. |
| Python GPT | dmind-aiservice-openai-gpt-python; retained but inactive in production-v2. |
| DMIND ASR | Active ECS/ECR artifacts exist; no matching source repository was found in the accessible aimet-tech repositories. |
| OpenAI ASR | Active ECS/ECR artifact exists; no matching source repository was found in the accessible repositories. |
| ECS converter | Active image and internal ALB exist. lambda-media-convert is a separate historical Lambda implementation; the active ECS converter source was not found. |
| Global ASR | global-asr-service maps to separate ECS, ECR, and S3 resources and is not referenced by the production-v2 controller task. |
Configuration
Task definitions provide queue URLs, bucket and database settings, the converter endpoint, Azure OpenAI endpoint and credential reference, on-prem ASR credentials, log groups, and the CloudWatch namespace.
Compatibility and deployment boundaries
The controller service version selects the question path, result format, and worker versions. Open the pipeline version routing →. ECS workers and the DSC Lambda alias are deployed independently. A controller release that emits a new request or result contract depends on compatible consumers, and in-flight FIFO messages retain the producer version's contract.