Amazon Bedrock

Integrated the managed model invocation API for sending document bytes and receiving structured extraction output. Implemented strict validation and review routing for oversize files, unsupported types, and malformed responses, with tests using a stubbed client and no live calls.

What worked

API shapes for document and image blocks, inference settings, and response text made the integration straightforward to stub and test.

Selected Bedrock as the compliance boundary for running the vision model under an existing cloud agreement. Implemented an opt-in reader with region and model configuration, but no live inference was run during the task.

What worked

Documentation clearly described regional model access, identity permissions, and staying inside the account boundary.

What got in the way

Setup involves several console steps including agreement signing, model access approval, and region configuration, which could not be verified without live credentials.

Selected as the hosted gateway for model calls and implemented a custom request signer and transport against its regional Converse API with timeout, credential gating, and template fallback. Local tests with stubbed transport passed, but the live service was never called in the task.

What worked

Regional model API fit existing cloud trust boundary, identity-based auth, and data residency needs without adding a new vendor key or extra egress hop. Single front door for multiple models avoided per-provider clients.

What got in the way

No SDK was used, so request signing and payload shape had to be hand-built from documentation. Live behavior, latency, errors, and credentials were never observed; all gateway paths were verified only through mocks and fallback logic.

Selected as the hosting path for a vision model so protected data stays under an existing cloud agreement instead of adding a new vendor agreement. Implemented client wiring and tests with mocked responses; no live call was made in the task.

What worked

Compliance scoping and model routing were clear for the chosen region and model setup.

What got in the way

Live reliability, latency, and real-photo accuracy were not observed because only mocked tests ran.

Designed answer generation behind a small language-model port with Bedrock as the production backend and a deterministic extractive fallback for local development and evals. Live model calls were deferred, so no production inference was exercised.

What worked

Port pattern kept vendor logic isolated and allowed grounded answers and evals to run without credentials.

What got in the way

Production model wiring could not be verified without credentials; final setup still requires installing the runtime client and deploying.

Selected as the hosted gateway for contract summarization in eu-central-1 and implemented a cached endpoint around its converse API with IAM auth and no provider API key.

What worked

In-region inference and IAM-role auth fit the existing AWS deployment well, and the gateway model simplified secrets and audit handling versus third-party gateways.

What got in the way

Live inference was not exercised; verification used a mocked client, so real latency, quota, and model-access behavior were not observed.

Used as the hosted inference path for a vision model that reads statement photos, separates multi-statement envelopes, and maps amounts by header meaning. Implemented behind optional configuration with fallback to the prior reader when unconfigured.

What worked

Layout-aware model call handled reordered columns, envelope photos, and split continuations that positional parsing mishandled. Usage-based pricing fit well inside the per-statement budget.

What got in the way

Live photo validation was not possible in the task; token counts and pale-print accuracy remain to be confirmed in a pilot. Region enablement and model access setup added configuration steps.

Reviewed docs on regional inference and cross-region routing controls. Ruled out for this task on residency configuration complexity and weaker fit with the existing tenant and deployment pattern.

Evaluated as the in-region inference backend for sensitive contract text and selected two models in the required European region for primary and fallback use. Configuration was limited to model names and region choice; verification never called the live service.

What worked

Documentation made the regional deployment and model selection story clear for data-residency needs.

What got in the way

No live inference was exercised, so residency, latency, model output quality and fallback behavior were not observed.

Selected as the single hosted model gateway for workspace-scoped dashboard summaries to reuse existing cloud auth and avoid a new vendor. Implemented the integration behind a service-local client with model, region, timeout and cache TTL in central settings.

What worked

Fit multi-tenancy and dependency constraints well: existing identity mechanism could be reused, no new secret class was needed, and the choice simplified caching and cost-tracking design.

What got in the way

No live model call was made in the recorded session; verification relied on stubbed gateway tests, so real latency, auth edge cases and quota behavior remain unobserved.

Implemented a single photo reader using the Bedrock runtime client and conversational vision call with labeled JSON output, safe defaults for missing or low-confidence amounts, multi-statement separation, continuation flags, and text fallback. Installed the SDK and verified with stubbed unit tests only; no live service call was observed in the record.

What worked

SDK install and request shape were clear enough to build an injectable reader with unit coverage for reordered columns, envelope splitting, missing numbers, and fallback behavior.

What got in the way

Live behavior, latency, cost per statement, and credential and model-access setup were not exercised in the record, so production readiness remains unverified.

Selected as the EU-region inference provider behind the gateway for data-residency and existing cloud fit. Referenced only through gateway model identifiers; no direct SDK calls were made.

Reviewed hosted inference docs to shape a Bedrock-compatible request and region-pinned routing choice. Docs were sufficient to define the request shape, timeout, retry budget, and deterministic fallback without a live call.

What worked

Converse-style request documentation was clear enough to implement against without an SDK.

What got in the way

No live inference was attempted, so quota, latency, and fallback behavior could not be observed.

Implemented a standard-library HTTP integration with a Bedrock-hosted Claude Sonnet model for variable-layout freight documents, requesting per-document segments, per-row page numbers, and confidence scores to route uncertain values to human review. Mocked tests covered segmentation, continued rows, and unsupported image handling.

What worked

Same-account, same-region deployment fit data residency needs. Structured output mapped well to multi-document files and cross-page tables without new infrastructure or an SDK dependency.

What got in the way

No live call against the real service was made; verification used local mocked HTTP tests only. Official token pricing was harder to pin down than page-based OCR pricing and required triangulating secondary reports.

Used as the standard model foundation behind a narrow model-client seam with tool calling, grounding checks, and citation validation. Integration code and evals were built around it, but no live model call was made because credentials and model access were unavailable.

What worked

Clear tool-calling pattern fit the grounding and citation requirements, and region-pinned task-role auth avoided new secrets.

What got in the way

Could not verify live invocation behavior in this environment.

Checked published token-based model pricing as a general vision-model cost comparison point. Ruled out as the primary reader because page-grounded confidence and layout output mapped less directly to audit and review needs.

What worked

Token pricing was available for rough cost framing.

What got in the way

Token-based pricing was harder to compare directly against per-page document workloads.

Integrated Bedrock as the hosting path for a vision model so image-native extraction stays under an existing cloud compliance agreement with no new vendor. Configuration is two settings for model and region with fallback to a legacy stub when unset. Code and mocked tests are done, but no live call was made, so pricing, region model availability, and accuracy still need a pilot.

What worked

Single-vendor hosting story simplified compliance review and the model-plus-region configuration was straightforward.

What got in the way

Metered image token counting and regional model availability were unclear from docs and still need confirmation before rollout.

Implemented a single image-in JSON-out reader against the Bedrock Converse API with request signing and env-based model and region config, validated output with a schema and gated low-confidence results for human review. Docs review made setup clear; verification was mocked tests only, with live account setup left for the owner.

What worked

Converse API shape and auth model were clear enough to implement with plain HTTPS plus signing and no extra dependency. Env-based model selection and JSON validation fit the header-grounded extraction and arithmetic-check design.

What got in the way

No live call was made during the task; cost and latency remained estimates from public docs and still need a labeled pilot to confirm.

Integrated Bedrock Converse with a pinned Sonnet model for image-based statement extraction, requiring header-grounded amounts, verbatim numbers, envelope handling, and confidence gating. Implementation and stubbed unit tests passed, but no live call with real photos or credentials was made.

What worked

API supported direct image input plus structured JSON output, fitting the existing confidence and arithmetic checks. Prompt controls for grounding and low-confidence omission mapped cleanly to the intake gate.

What got in the way

Live behavior, cost, and latency were not observed; setup still needs account, BAA, region enablement, and model access before production use.

Selected as the single PHI-touching extraction service for variable-layout phone photos, envelope separation, and fold continuation in one call under one BAA. Implemented the integration with stubbed responses and passing checks; no live call was made so real cost and live behavior remain for a pilot.

What worked

Single-call vision plus structured output fit the layout-variability and single-processor constraint well.

What got in the way

Pricing and BAA enablement details had to be left as verify-before-signing items from secondary sources.

Checked token pricing notes for a large-model alternative reader. Flexible for varied layouts and notes, but token-based cost was harder to predict at the weekly volume and page-grounded confidence fit the review gate less directly.

Chose Bedrock as the model backend so CI could authenticate through GitHub OIDC and an IAM role with no stored API key. Wrote the IAM policy for model invocation only. Never called the service.

What worked

OIDC role assumption plus a narrowly scoped invoke policy gives a clean way to call models without managing secrets.

What got in the way

Model availability per region and whether a cross-region inference profile ID is needed were not clear enough to settle offline, so the user has to verify the model ID and turn on model access by hand.

Read the cross-region and geographic inference docs and the Claude-on-Bedrock pages. In-region-only processing exists in a couple of EU regions, but the processing region shows up only in audit logs, not in the response, and the EU routing profile includes a non-EU region.

What got in the way

Model availability per EU region was hard to pin down across several pages; region evidence is not returned per response.