AI Engineer

Decode Talent is partnering with a well-established, vertically integrated direct-to-consumer brand that designs, sells, and manufactures made-to-order products. They’re making a significant, C-suite-backed investment in AI across sales, customer support, and manufacturing, and they’re hiring their first dedicated AI Engineer to lead the build.

The Opportunity

You’ll build the company’s AI applications: assistants that answer customer and associate questions, tools that search a complex product catalog and internal knowledge, and workflows that take routine work off people’s plates. You’ll work with large language models (Claude on Amazon Bedrock or Microsoft Foundry to start) and ship them into live, business-critical systems.

This is a new discipline for the business. The product catalog is highly configurable, knowledge lives in many places, and core systems are legacy. You’ll decide what to build first, prove it works, and set the patterns every future use case follows. This role is for a hands-on, high-agency builder who designs the model layer, retrieval, and evals, not someone who only uses AI tools out of the box.

LLM Applications

• Design, build, and ship production applications on Bedrock, Foundry, and Claude, including prompt design, tool use, and multi-step agent workflows
• Build agents and tool layers with MCP (servers, tool design, auth) and frameworks such as Bedrock AgentCore, Strands Agents, or LangGraph
• Route across models and providers (Bedrock, Foundry, OpenRouter, LiteLLM) with fallbacks, choosing models on quality, latency, and cost
• Control cost and latency with prompt caching, semantic caching, rate limits, and token budgets
• Implement guardrails for accurate, on-brand, safe outputs, including PII handling and audit trails

Retrieval & Data

• Build RAG pipelines over product, policy, and support content: ingestion, chunking, embeddings, retrieval, and re-ranking
• Build hybrid (dense + sparse) retrieval with reranking, and measure retrieval quality rather than assume it
• Select and tune vector databases (OpenSearch, pgvector, Pinecone) for search over a configurable catalog
• Enforce permission-aware retrieval
• Partner with Data & Analytics on a Snowflake/dbt stack and MySQL application databases

Evaluation & Operations

• Build eval harnesses: golden datasets, regression tests, LLM-as-judge, and human review; measure grounding and hallucination
• Instrument systems with tracing and per-agent telemetry (Langfuse, OpenTelemetry) and own production quality, latency, cost, and failures
• Serve AI features behind documented internal services with versioning, rollback, and fallbacks

Technical Leadership

• Establish reusable patterns for prompt management, evaluation, observability, and deployment, making pragmatic build-vs-buy calls
• Set shared standards with Data, Analytics, and ML teams
• Mentor engineers and communicate clearly with technical and executive audiences

What You Bring

• 5+ years of software engineering, with 2+ years building production LLM applications
• Hands-on experience with Bedrock, Foundry, Claude, or comparable foundation model platforms
• Experience running LLM systems beyond a single vendor SDK: gateway/router, fallbacks, caching, cost controls
• End-to-end RAG knowledge: embeddings, chunking, hybrid retrieval, reranking, evaluation
• Production experience with a vector database or vector search
• Hands-on MCP experience and agent/tool-use patterns
• Production LLM observability and evals: tracing, regression sets, grounding checks, testing non-deterministic systems
• Strong Python and solid SQL, comfortable reading a live transactional schema (MySQL preferred)
• Technical leadership without direct reports, plus modern SDLC practices (Git, CI/CD, code review)
• Excellent communication, translating technical detail into business impact
• High agency and comfort with ambiguity

Nice to Have

• AWS (Lambda, API Gateway, S3, IAM, CloudWatch, OpenSearch) and infrastructure as code
• LangChain, LlamaIndex, Bedrock Agents/Knowledge Bases, Ragas, or DeepEval
• Fine-tuning or self-hosted open-weight models (LoRA, vLLM), and judgment on when they’re worth it
• Modern data stack experience (Snowflake, dbt, Dagster/Airflow, Fivetran)
• Experience working alongside a legacy PHP/MySQL application
• Background in e-commerce, retail, manufacturing, or DTC
• Daily use of AI coding tools (Claude Code, Codex, Cursor), and experience standing up a new technical function in a non-tech-native company