We are looking for senior generalist engineers with real depth in LLM and agentic systems. This is not a ticket execution role: you get the problem and the context, you propose the solution, you build it, you ship it, and you defend the technical decisions in front of the client.
The Role: What You Are Actually Doing
You will take a production-grade LLM agent from read-only insight toward supervised action: hardening it for scale and staging it up a capability ladder (Explains 12 Recommends 12 Orchestrates 12 Acts). That means agent orchestration, evaluation you can trust, observability, cost and latency engineering, resilience, and safe write-actions with a human in the loop, all on AWS. Fluency with AI-assisted engineering (Claude Code, Cursor, Copilot, or equivalent) is the baseline here, not a differentiator - but you sign the code, and "the AI wrote it" is never an answer when something breaks in production.
Key Responsibilities
- Own End-to-End Delivery: Take a production LLM agent from requirement to production deploy, and defend the architecture and trade-offs directly with the client.
- Agent Orchestration: Build and harden orchestration (LangGraph / LangChain or equivalent) - routing, tool-calling, planning, synthesis, and state management.
- Tool & Retrieval Integration: Integrate tools over MCP and keep a growing tool surface fast and correct, including BM25, hybrid, or vector retrieval as scale demands.
- AWS Bedrock & AgentCore: Run models on Bedrock and Bedrock AgentCore - model selection/routing, guardrails, memory, and regional residency profiles.
- Evaluation & Observability: Build the eval harness (golden sets, LLM-as-judge, quality gates wired into CI) and instrument the system with OpenTelemetry for per-session token, cost, and latency attribution.
- Cost, Latency & Resilience: Drive down cost and latency with real levers (model routing, prompt caching, payload pruning, parallelizing independent calls) behind a regression gate, and build in circuit breakers, fallbacks, and dead-letter handling.
- Safe Write-Actions: Design and stage write-actions with least-privilege permissions, human-in-the-loop approval, plan versioning, audit trail, and rollback - released behind feature flags to a small cohort first.
What We Are Looking For
- 5+ Years in Software Engineering: At least 2 of them genuinely at a senior level, with strong Python in production and comfort picking up TypeScript or Go when a project calls for it.
- Hands-On LLM Application Engineering: Prompt design, tool/function calling, structured output, context management, and token budgeting, in a system real users hit.
- Agent Orchestration Experience: Built or operated orchestration with a framework like LangGraph or LangChain, or hand-rolled, beyond single-prompt calls.
- Managed LLM/Agent Platform in Production: AWS Bedrock, Google Vertex AI, or Azure AI Foundry - model invocation, streaming, guardrails, and agent tooling. We use Bedrock and Bedrock AgentCore; equivalent depth on Vertex AI or Azure transfers directly.
- Evaluation, Retrieval & Observability: Eval harnesses and golden/reference sets, vector or hybrid search in production, and OpenTelemetry-based distributed tracing with token/cost/latency attribution.
- Production AWS, CI/CD & IaC: Real IAM, networking, storage, and observability experience, CI/CD pipelines versioned as code, and Terraform in production.
- Working English: Comfortable defending system design and technical decisions directly on client calls.
Our Tech Stack
AWS Bedrock, Bedrock AgentCore, LangGraph/LangChain, MCP, Terraform, GitHub Actions/GitLab CI, OpenTelemetry
