Job Summary
You will be a core engineer on the AI intelligence layer of an enterprise-grade AI application. Your day-to-day work spans designing and building LLM-powered pipelines, agentic reasoning workflows, semantic retrieval systems, and knowledge graph integrations that serve real business users at scale. You write Python every day. You are responsible not just for building AI features but for ensuring they work reliably and measurably in production — not just in demos.
Key Responsibilities
-
Design and build natural language to SQL (NL2SQL) pipelines — schema grounding techniques, few-shot prompt construction, SQL validation and sanitisation, guardrail enforcement, query explanation generation, and robust error recovery for malformed or ambiguous queries
-
Build and maintain LangGraph-based agentic workflows — state machine design for multi-step reasoning, tool-use orchestration, memory and context management, human-in-the-loop interruption patterns, and agent output validation
-
Implement retrieval-augmented generation (RAG) and graph-enhanced RAG (GraphRAG) systems — chunking strategies, embedding model selection and evaluation, hybrid retrieval (dense + sparse), reranking, entity traversal over graph databases, and grounded answer synthesis
-
Design and maintain knowledge graph schemas for business domain semantics — node and relationship type definitions, traversal patterns, semantic similarity scoring, and integration with vector search for combined retrieval
-
Build and operate LLM evaluation pipelines — RAGAS metrics (faithfulness, answer relevance, context recall), LLM-as-judge patterns, custom domain-specific quality metrics, regression testing on prompt changes, and hallucination rate monitoring in production
-
Optimise LLM API usage across Vertex AI and Gemini — prompt caching, model tiering (Flash vs Pro selection criteria), context compression, cost attribution per feature, quota management, and graceful fallback strategies
-
Implement vector embedding pipelines end-to-end — document chunking, embedding generation (text-embedding-gecko or equivalent), index construction and maintenance in AlloyDB pgvector and Vertex AI Vector Search, and scheduled freshness updates as source data changes
-
Write clean, async FastAPI endpoints that expose AI capabilities to frontend and downstream services, with proper streaming support, error handling, and OpenAPI documentation
Skill Requirements
-
Python — 5+ years of production-grade engineering: async / await, Pydantic, type hints, modular package design, pytest, and clean separation of concerns; not just notebook scripting
-
LangChain or LangGraph — production experience building chains, agents, or state machines that handle real-world edge cases and fail gracefully; demonstrable shipped systems, not tutorials
-
LLM APIs — Vertex AI / Gemini, OpenAI, or Anthropic; deep practical understanding of context windows, token budgets, structured output enforcement (JSON schema, function calling), and cost optimisation strategies
-
RAG architectures — chunking strategies, embedding model evaluation, retrieval quality measurement (precision, recall, MRR), reranking approaches, and hybrid search implementation
-
SQL — fluent in complex query writing: joins, window functions, CTEs, subqueries; must be comfortable reading and writing BigQuery and PostgreSQL dialect SQL without assistance
-
Vector search — hands-on production experience with at least one: pgvector, Vertex AI Vector Search, Pinecone, Weaviate, or ChromaDB; including index configuration, ANN parameter tuning, and embedding dimension trade-offs
-
Prompt engineering — systematic, evidence-based approach: few-shot design, chain-of-thought prompting, output structuring, negative example inclusion, prompt versioning, and iterative evaluation methodology
Other Requirements
-
GraphRAG or knowledge graph experience — Neo4j, Spanner Graph, or Amazon Neptune with GQL or Cypher; entity-relationship traversal for context-enriched retrieval
-
NL2SQL specialisation — prior work on text-to-SQL systems, schema linking, value grounding, or SQL correctness evaluation benchmarks (Spider, BIRD)
-
LLM fine-tuning — instruction tuning, LoRA or QLoRA on domain datasets; experience setting up fine-tuning pipelines and evaluating fine-tuned models against base models
-
FastAPI production patterns — background tasks, lifespan event management, dependency injection, streaming responses with SSE, and middleware design
-
GCP Vertex AI platform beyond inference — Vertex AI Pipelines, Model Registry, and Evaluation Service
-
Any enterprise domain depth — retail analytics, financial services, healthcare, supply chain, or HR analytics; domain knowledge accelerates prompt design and evaluation