Job Description

Job Summary:

Key Responsibilities

Build and extend PDF text & layout extraction using PyMuPDF for text-based PDFs, and Claude Vision for scanned, multi-column, or complex pages, producing bounding-box-anchored regions.
Generate multilingual/cross-lingual embeddings (via Bedrock) to shortlist candidate PDF regions matching each English instrument item.
Integrate a vision-capable Claude model on Amazon Bedrock as a verifier to confirm the correct source region for each item, with confidence scoring and a small-model fallback chain; ensure unmatched or low-confidence items are flagged rather than guessed.
Build deterministic placement logic (AWS Lambda) that clones the English skeleton, injects matched text, and validates structural integrity — string IDs, ordering, and formatting tags.
Design and build the agent orchestration layer using AWS Agent Core / MCP Server, connecting extraction, retrieval, verification, and placement as distinct tools with clear interfaces.
Implement the extract retrieve verify place validate workflow and guardrails using AWS Strands (Agentic SDK), ensuring structure-locked output, zero invented text, and full source traceability.
Support the linguist review & approval tool, showing each output string alongside its PDF source region, and implement a per-string audit trail written to Amazon S3 / DynamoDB.
Integrate with the existing eCOA ingest / screenshot-review pipeline: accept English instrument JSON and translated PDFs as inputs, and emit a draft JSON in the required target schema.
Run validation across sample instruments and priority locales (including at least one non-Latin-script locale), and track/report placement accuracy and flag-rate metrics (target 90% correct placement, zero invented translations).
Participate in daily scrum and weekly governance calls with Sonata and Medidata stakeholders; track progress in Jira.

Required Skills & Experience

Hands-on experience with AWS Bedrock — invoking foundation models (Claude), including vision-capable models for document/image understanding.
Experience building agentic AI systems using AWS Agent Core / MCP Server (Model Context Protocol) and AWS Strands, or comparable agent orchestration frameworks (e.g., LangGraph, LangChain agents).
Strong Python development skills, including experience with PDF parsing libraries (PyMuPDF, pdfplumber, or similar).
Experience with embeddings-based retrieval / semantic search, including cross-lingual or multilingual embeddings.
Working experience with AWS Lambda, Amazon S3, DynamoDB, and CloudWatch.
Solid understanding of prompt engineering and LLM verification/guardrail patterns — confidence thresholds, hallucination avoidance, fallback chains.
Comfortable working with structured JSON schemas and schema validation.
Experience delivering in Agile/Scrum teams using Jira.

Preferred / Nice to Have

Experience in healthcare, life sciences, or clinical trials technology (eCOA, ePRO, EDC systems).
Exposure to localization/translation-adjacent workflows or NLP/linguistics-adjacent projects.
Experience with OCR or vision-based document understanding at scale.
Prior experience integrating AI/automation solutions with third-party enterprise platforms via APIs.