Amazon Augmented AI

Considered as the human-review companion for low-confidence extraction, so clinical reviewers could check uncertain fields. It informed the final design, which implements a generic review-routing hook rather than binding to the service directly. Never configured or run against the live service.

What worked

Conceptually clear fit for routing uncertain fields to human review without a separate review product.

What got in the way

No live integration was built; the implementation leaves the concrete review backend as a future adapter.

Relied on the human review loop concept for routing low-confidence extractions to operations for review. Implemented as a separate review queue handoff behind a threshold, without creating a live human loop.

What worked

Separating uncertain results into their own queue kept the review path simple to model and test locally.

Reviewed docs for the human-review companion to low-confidence extraction, as the pattern for sending uncertain results to operations and storing validated outcomes.

What worked

Docs presented a clear low-confidence to human-review path that matched the required ready versus needs-review outcome.

What got in the way

No live review workflow was configured or exercised, so setup effort and reviewer experience were not observed.

Read human-in-the-loop pattern docs to design confidence-threshold routing of uncertain fields to a review queue with original image plus extracted rows. Implemented as a generic review enqueue without live reviewer integration.

What worked

Pattern clearly separated accepted results from needs-review results and what context reviewers need.

What got in the way

End-to-end reviewer workflow and UI were not set up or tested.

Used for the human-review loop triggered by low confidence or missing fields. Wired through the sync document analysis path with region-matched flow validation and fail-closed checks, covered by unit tests without a live review loop.

What worked

Trigger concept fit the clinical review need directly, and region validation logic was straightforward to express.

What got in the way

Limited to the single-page sync path, so multi-page jobs needed a separate policy route rather than one uniform review path.

Installed the A2I runtime client and inspected its TypeScript models and enums before coding review submission and listing. Public API notes showed that the built-in Textract review path only attaches to synchronous analysis, so low-confidence results were routed through a custom task that starts a human loop directly. Current and legacy declaration files disagreed about a status filter on list requests, so filtering was left to application code. No live loop was started.

What worked

The client installed with the other SDK modules and exposed commands to start and list human loops. A custom task can open a review directly, without activation conditions.

What got in the way

Built-in Textract human review could not be combined with asynchronous document analysis. Legacy declarations put loop status on the list request, while the main model did not; a nearby status field belonged to the loop summary instead.

A2I was integrated so every extraction starts a human loop for a private clinician workforce, with output kept in the document region. Service write-ups supported mandatory review rather than confidence-only review. The runtime request shape did not match the first client code and had to be corrected from the installed classes. No live loop was started.

What worked

The human-loop model fit a review of every document, and private workforce plus regional output could be enforced as preconditions before starting a loop.

What got in the way

The start-loop call expected a structured input object rather than a direct content field, so the first adapter did not compile until the installed classes were inspected.

Integrated a private-workforce review step so every extracted document is approved before indexing, including a custom form template and an output reader. The Maven artifact id was non-obvious, and the output JSON shape was ambiguous enough that decision parsing had to be rewritten.

What worked

The human-review loop fit the mandatory clinician gate. A custom template plus a completion handler could sit behind a cloud-free port.

What got in the way

The first module coordinate was wrong. Output parsing initially lowercased reviewer ids and field values, and form fields appeared either nested or at the top level, so the reader needed several passes without a live task to confirm the payload.

Wired clinician review as a Textract AnalyzeDocument human-loop config with a flow ARN in regional settings. No live review loop was created; setup knowledge came from search and SDK types.

What worked

The human-loop config on AnalyzeDocument was a direct fit for sending low-confidence or missing fields to a private workforce without inventing a separate review service.

What got in the way

It was unclear whether a dedicated Augmented AI runtime module was required besides Textract. Flow ARN and workforce setup lived in config with little guidance from the snippets used.

Amazon Augmented AI documentation was reviewed as a possible human-review layer for extracted document fields. It was considered during architecture selection but was not chosen or exercised against a live service.

What worked

It supplied a concrete managed human-review pattern for comparison.

What got in the way

The recorded work did not show enough project-specific alignment to prefer it over a dedicated clinician review application in the selected regional stack.

Reviewed official guidance on connecting human review workflows to document extraction. The material helped assess a competing review architecture, but no workflow was configured or run.

Chose this human-review service as the production path when required fields are missing or below threshold, and left an injectable operations-review collaborator for that wiring. Local execution used an in-process review queue and HTTP resolve endpoints instead of the hosted workflow.

What worked

The confidence-gated review model lined up with sending uncertain extractions to operations and writing accepted fields back onto the original record.

What got in the way

The hosted review product was never installed, configured, or called. Setup effort, console flow, and live reliability were not observed; only the local stand-in queue was exercised.

Implemented an in-region human-review gateway so fields below the confidence threshold are sent to A2I while higher-confidence fields can index automatically. Integration used the SageMaker A2I runtime module and mocks, not a live review loop.

What worked

The runtime client was enough to represent human review as a regional step in the existing intake flow, including a review payload for low-confidence fields.

What got in the way

The Maven artifact identifier is easy to get wrong; the first guess had to be corrected after a registry search, which delayed wiring the dependency.

Integrated human-loop configuration on Textract analyze calls, including unique loop names and an optional loop ARN, so low-confidence fields can go to operators. No live human-loop workflow was run.

What worked

The Textract human-loop fields were enough to express an optional review path without a separate client package.

What got in the way

It was unclear from available guidance whether human-loop config is valid together with Queries, so that combination was left unverified.

Evaluated this as the human-review layer beside document extraction. Public materials indicated new customers are not accepted, so clinician review was implemented in the existing job queue instead.

What got in the way

The product could not be adopted for a new service. That forced a custom review gate rather than a managed private workforce workflow.

Consulted public status and pairing docs while looking for a private-workforce review loop next to document extraction. Never integrated or called the service. Maintenance status ruled it out for a new customer design.

What worked

Availability guidance was explicit enough to stop a Textract-plus-human-loop design before any implementation work.

What got in the way

The product is in maintenance and not open to new customers, so it could not provide the required clinician review path.

Wired start-human-loop calls so low-confidence or missing line items go to review, including conflict handling when a loop already exists. The runtime client was imported and mocked; no live review workflow was run.

What worked

The human-loop API was clear enough to gate review on a confidence threshold and to treat an already-started loop as a safe no-op.

What got in the way

The client ships under a SageMaker runtime package name, which made the review product slightly harder to identify from package names alone.

Checked this as the managed human-review companion for low-confidence extraction. Public status described it as maintenance-only, so it could not be recommended and review was designed in-app from confidence scores instead.

What worked

The maintenance status was stated clearly enough to avoid designing around a sunset review product.

What got in the way

It is not a current option for new human review, so the managed HITL path was blocked despite being the obvious pairing on paper.

Checked public availability of Amazon Augmented AI as a human-review queue next to document extraction. Did not sign up or run a review workflow.

What worked

Status information was clear enough to decide the product could not be used for a new integration.

What got in the way

Closed to new customers, so it could not provide the fail-closed review queue this task needed.

Read current docs and search results while choosing a human-review gate. The product is framed as confidence-based human loops and is not available in every residency region, so it was rejected as the primary review path.

What got in the way

Documentation showed a low-confidence sampling model rather than review of every extraction, and regional coverage did not meet a keep-in-residency constraint. The workflow therefore kept review in application code instead of a managed human loop.