Job description:

KEY RESPONSIBILITIES

· Design and own multi-stage ingestion pipelines — handling HTML, PDF, and image sources with layout parsing, metadata extraction, and vector storage

· Architect RAG systems with hybrid search (BM25 + semantic), document versioning, and cross-reference resolution

· Build production-grade FastAPI services with typed response envelopes, OpenAPI compliance, and Langfuse tracing integration

· Engineer prompt systems — structured prompts, prompt versioning, few-shot strategies, and judge-based evaluation

· Integrate and manage LLM routing via LiteLLM: model fallback, cost control, and per-route configuration

· Design agentic workflows using LangGraph: multi-step retrieval, tool use, and conditional branching

· Build and maintain knowledge graphs in NebulaGraph / Neo4j — entity extraction, relationship modelling, and domain ontology alignment

· Implement graph-augmented retrieval (GraphRAG) — combining vector search with graph traversal to surface contextually connected information beyond chunk-level retrieval

· Own entity linking and co-reference resolution pipelines that connect ingested documents to graph nodes

· Lead RAG evaluation initiatives — define metrics, build eval datasets, and run regression cycles

· Drive observability standards — tracing, cost attribution, and latency profiling via Langfuse

· Collaborate on K8s deployment patterns for AI services: resource limits, GPU scheduling, and health probes

· Mentor junior developers and conduct code and prompt reviews

REQUIRED SKILLS

· Python (5+ years) — async, concurrency patterns, production packaging

· Deep understanding of RAG — hybrid retrieval, reranking, chunking strategies, embedding model selection

· FastAPI — dependency injection, middleware, background tasks, async patterns

· Prompt engineering — structured prompting, chain-of-thought, evaluation-driven iteration

· LLM API integration — OpenAI-compatible APIs, AWS Bedrock, or similar

· Vector DB expertise — Weaviate or equivalent: schema design, indexing, and filtering

· Document parsing at scale — Docling, layout models, VLM-based extraction from PDFs, HTML, and images

· Graph DB — NebulaGraph or Neo4j: schema design, Cypher / nGQL queries, knowledge graph construction

· Observability mindset — tracing, evaluation loops, cost-aware system design