Overview

We're building agentic AI systems that autonomously reason, plan, and act across complex workflows.

Responsibilities

• Design and build AI agents — autonomous systems that use tools, browse the web, call APIs, manage memory, and complete multi-step tasks with minimal human intervention
• Architect agentic workflows — orchestrate multi-agent pipelines (planner → executor → verifier patterns), handle tool-use loops, HITL checkpoints, retries, and failure recovery
• Integrate LLM providers — work with Anthropic Claude, OpenAI, Gemini; implement prompt caching, structured outputs, and context management at scale; select the right model for the right task
• Build knowledge pipelines — design RAG systems, hybrid retrieval, vector stores, embedding pipelines, and memory layers that give agents contextual awareness
• Own backend reliability — REST/GraphQL APIs, async job queues, observability, latency optimization, cost tracking for token usage
• Drive system design decisions — lead architecture discussions, define component boundaries, evaluate build vs. buy tradeoffs, and document decisions (ADRs, diagrams)
• Maintain CI/CD pipelines — keep deployments fast, safe, and automated across dev/staging/prod environments
• Evaluate agent behavior — build LLM tracing, regression frameworks, and evals pipelines to detect behavioral drift, intent failures, and quality degradation across model updates
• Stay ahead of the curve — evaluate emerging standards (MCP, A2A, reasoning models, new frameworks) and make adoption recommendations

Qualifications

• Deep experience with at least one major LLM provider SDK (Anthropic, OpenAI, Gemini)
• Prompt engineering depth: few-shot design, chain-of-thought, structured outputs, tool/function calling, multi-turn context management, prompt caching, token budgeting
• Built real production agents — not just chatbots; systems that take actions, use tools, and operate autonomously
• Understands agent patterns: ReAct, plan-and-execute, multi-agent orchestration, reflection loops, supervisor/worker hierarchies
• Familiar with MCP (Model Context Protocol), A2A, and agent frameworks (LangGraph, CrewAI, AutoGen) — and when NOT to use them
• Knows how to handle non-determinism: retry strategies, output validation, graceful degradation
• Reasoning models — knows when to use o1/o3, Claude Extended Thinking vs. standard models; treats model selection as an architectural decision, not an afterthought
• Model routing & cascading — routes queries to right-sized models (Haiku/Flash for simple, Opus/GPT-4o for complex) based on cost/capability tradeoffs; doesn't default to the largest model for everything
• Compound AI systems — combine multiple specialized models and tools in pipelines rather than relying on a single general-purpose model
• Human-in-the-loop (HITL) — designs checkpointing for long-running agents, approval workflows, interrupt/resume patterns; knows when and where agents need human gates
• Persistent agent state — durable execution patterns (LangGraph checkpointers, Temporal-style workflows); agents that survive restarts and partial failures
• Prompt injection & AI security — understands attack surfaces unique to tool-using agents; builds with guardrails in mind (NeMo Guardrails, Llama Guard, output schemas)
• Semantic caching & batch inference — reduces LLM costs through response-level semantic caching and async batch APIs; thinks about cost at 10x before it becomes a problem
• LLM-specific observability: Langfuse, LangSmith, Helicone, or Arize Phoenix — prompt/response tracing, token cost per flow, per-step latency, session replay
• Evals methodology: RAGAS, DeepEval, Promptfoo, or Braintrust — systematic evaluation of LLM outputs, not vibes-based testing
• Treats prompts as code: versioning, A/B testing, rollback; understands how prompt changes ripple into downstream agent behavior
• Knows how to run AI/non-deterministic systems in CI — mock vs. live modes, cost controls, behavioral regression (not just output format matching)
• Cost tracking and optimization: token usage per workflow, model spend dashboards, alerting on cost regressions
• RAG pipeline depth: document ingestion, chunking strategies (fixed, semantic, agentic), embedding, retrieval, reranking
• Hybrid search — dense (vector) + sparse (BM25/keyword) retrieval; understands where pure vector search underperforms and how to combine approaches
• Awareness of GraphRAG and contextual retrieval (prepending context summaries to chunks) for structured knowledge domains
• Worked with vector databases: Pinecone, Weaviate, pgvector, or similar
• Understands knowledge graph concepts and when structured retrieval outperforms embedding-based approaches
• Can design distributed systems end-to-end: from data flow diagrams to API contracts to deployment topology
• Knowledge & Retrieval
• Knows when to use async vs. sync, event-driven vs. request-response, monolith vs. microservices — and can justify the choice
• Comfortable with architecture patterns: event sourcing, CQRS, saga pattern for multi-step workflows, circuit breakers
• Documents decisions clearly (ADRs, architecture diagrams, onboarding docs) — not just builds, but communicates design
• Has opinions on API versioning, backward compatibility, and service contracts
• Thinks about scalability and cost together — not just 'will it scale' but 'what does it cost at 10x'