Senior RAG - Document AI Engineer

Chryselys; Hyderabad, Chennai District; Hybrid; INR 300,000 - 600,000; Full time

Job summary

Chryselys in Hyderabad, India, is hiring a Senior RAG - Document AI Engineer to design, build and operate Python/FastAPI microservices that convert heterogeneous enterprise documents into accurate, cited answers. You own retrieval quality and the evidence behind it.

Responsibilities include format-aware parsing, a configurable chunking strategy, LLM-based metadata extraction, a robust retrieval stack, and an evaluation harness.

Qualifications

• 6–10 years software engineering experience
• 2+ years building retrieval or document-AI systems in production
• Production RAG at scale with large corpora
• Latency targets and operational maturity in live systems
• Chunking as a design decision for documents
• Knowledge of hybrid dense + sparse retrieval and evaluation measures

Responsibilities

• Build document processing services for text, OCR, and charts across formats
• Create a configurable chunking service with versioned profiles
• Implement LLM-based metadata extraction with strict JSON schemas
• Develop the retrieval service: query rewriting, density and BM25, rank fusion, reranking
• Own the evaluation harness and gating releases by quality metrics
• Instrument, monitor and support services in production

Skills

RAG at scale

Python

FastAPI

Distributed systems

CI/CD

Retrieval systems

Tools

Docker

AWS

OpenSearch

Bedrock

Textract

Job description

Senior RAG - Document AI Engineer

Design, build and operate the Python/FastAPI microservices that turn heterogeneous enterprise documents into accurate, cited answers from document parsing and chunking through enrichment, hybrid retrieval and evaluation. You own retrieval quality and the evidence for it.

Responsibilities

• Build document processing services: format-aware parsing across native text, OCR and vision-language models; table and figure extraction.


• Build a configurable chunking service with strategies selectable per document class, maintained as versioned pipeline profiles.


• Implement LLM-based metadata extraction into strict JSON schemas, with confidence scoring and human-in-the-loop review.


• Build the retrieval service: query rewriting, hybrid dense + BM25 with rank fusion, metadata pre-filtering, cross-encoder reranking, context assembly with citations.


• Own the evaluation harness — golden Q&A sets, recall@k, nDCG/MRR, groundedness — and gate releases on measured quality.


• Instrument, monitor and support the services in production.

Qualifications

• 6–10 years software engineering, with 2+ years building retrieval or document-AI systems used by real users in production.


• Production RAG at scale. 100k+ documents and millions of chunks; p95 query latency under 2s, sustained under concurrent load.


• Operational maturity. Incremental and delta ingest; has re-indexed a live corpus without downtime after a chunking or embedding-model change.


• Chunking as a measured design decision. Hierarchical parent–child and section-aware strategies — not a single global token size.


• Retrieval depth. Hybrid dense + sparse (BM25) retrieval and rank fusion; cross-encoder reranking; embedding model selection and evaluation.


• Retrieval evaluation in practice. Recall@k, nDCG/MRR and a groundedness measure, used to justify changes.


• Document processing. Messy real-world PDF, PPTX and DOCX; OCR pipelines and their failure modes; vision-language models for charts and infographics.


• Python and FastAPI in production. Python 3.11+, async, Pydantic, streaming/SSE, OpenAPI contracts, API versioning.


• Service engineering. Queue and worker patterns, idempotency, retries, dead-letter handling; Docker, pytest, Git and CI; structured logging and tracing.


• AWS as a consumer. S3, ECS/EKS, Lambda, Bedrock, Textract, OpenSearch; a vector database in production with collection design and metadata filtering.

Preferred Qualifications

• Life sciences, pharma or other regulated content; PII/PHI handling.


• Multimodal retrieval; MCP or agent tool surfaces.


• Judgement on frameworks — has used LangChain or LlamaIndex and can say when not to.