AI Platform Pipeline

Responsibility

dmind-aiservice-aicontroller owns the prediction lifecycle used by Prescreening and IVR. It accepts text, audio, video, or choice answers; coordinates conversion, transcription, depression scoring, and GPT classification; persists partial state; and exposes a single polled result.

Production-v2 is currently running controller image dmind-ai-service-ai-controller:v1.2.0 in AWS account 564141170168, region ap-southeast-1.

End-to-end flow

Media conversion is synchronous from the controller's perspective: it uploads the source, calls the converter, and enqueues the converted key only after conversion completes. The ASR implementations are competing consumers of one FIFO; the controller does not select an implementation for each message.

Public and worker-facing APIs

Route	Role
POST /prediction	Creates a versioned prediction with interview_id, interview_source, and method type.
POST /submit-answer	Accepts one text/media answer. Media is stored and transcribed; text can immediately unlock downstream tasks.
POST /submit-multiple-answers	Accepts the choice-method answers and performs configured score/action calculation.
GET /result?prediction_id=...	Returns PROCESSING until every required result for the selected method is complete, then returns FINISH.
POST /predictions/text	Internal callback for a completed transcript.
POST /predictions/result	Internal callback for DSC/GPT result fragments.
GET /question-path	Returns the configured question path for a service version and method.

The production workers normally use the result FIFO queue instead of calling the internal callback routes directly. The controller consumer reads message attribute type and dispatches store_text or save_result.

Although request binding permits an omitted type, clients using result format 1_1_0 or 1_2_0 require TEXT, AUDIO, VIDEO, or CHOICE; without a method the prediction cannot reach FINISH. CHOICE routes, bundle entries, and result fields remain for compatibility, but the current Prescreening and IVR journeys do not select that method.

Version routing

The requested service version is a compatibility bundle, not only a response label.

Service version	Question path	Result format	ASR	OpenAI	DSC	Choice calculation
1_0_0	1_0_0	1_0_0	1_0_0	1_0_0	1_0_0	none
1_0_1	1_0_0	1_0_0	1_0_1	1_0_0	1_0_0	none
1_1_0	1_1_0	1_1_0	1_0_1	1_0_0	1_0_0	1_0_0
1_2_0	1_1_0	1_2_0	1_0_1	1_0_0	1_0_0	1_0_0

Changing a dependency mapping changes existing clients that request that service version; an incompatible mapping is represented by a new bundle.

Task triggering

After a transcript is stored, the controller checks whether all answers required by each task have TRANSCRIBE_SUCCESS status.

Task	Required input	Output
openai_two_q_plus_sad	qs-01-02	sadness flag and sad chief-complaint terms
openai_two_q_plus_boredom	qs-03, qs-03-04-01, qs-03-04-02	boredom flag
openai_two_q_plus_suicidal_thought	qs-07-01	suicidal-thought flag
openai_suicidal_severity	qs-07-03	suicidal severity, rank, and detail
dsc_depression_rank	all configured narrative answers	depression rank
choice_calculation	configured qsc-* choices	two-question flags, depression.rank_choice, depression.level, and suicidal.rank for configured severe choices; it does not populate suicidal.severity

The result format derives depression level from configured outputs and reports FINISH only after every method-required result is successful.

Active versus retained implementations

• The Go worker dmind-aiservice-openai-gpt-golang is the current production-v2 GPT consumer.
• The production-v2 Python GPT service definitions are retained but not running; dmind-aiservice-openai-gpt-python remains a compatibility-era implementation.
• The current converter is an ECS service behind an internal ALB. lambda-media-convert is the earlier Lambda implementation and is not in this runtime path.
• global-asr-service is a separate service. The controller production-v2 path uses the DMIND-specific ASR FIFO and its competing workers.