Architect

We’re hiring an AI Architect to design and govern the end-to-end architecture for AI solutions spanning multi-agent orchestration, knowledge-graph reasoning, retrieval-augmented generation (RAG), and evaluation at scale. You will define model strategy, reasoning patterns, data interfaces, and safety/guardrails; establish reusable blueprints; and partner with engineering, security, and data teams to deliver reliable, explainable, and cost-effective AI capabilities. Key ResponsibilitiesAI Architecture & Patterns: Define reference architectures for LLM-powered systems (planner/reasoner, router/composer, tool/connector mesh), including typed I/O contracts and failure isolation.Knowledge & Retrieval Strategy: Architect coexistence of Knowledge Graph reasoning and RAG baselines; select embedding/vector stores, chunking, retrieval/reranking, and subgraph/query patterns.Model Strategy & Routing: Establish model selection policies (small→large escalation), prompt/adapter patterns, caching, and cost/latency budgets; document routing outcomes.Evaluation & Quality Gates: Design test sets and scoring rubrics (faithfulness/correctness, precision/recall, multi-hop coverage, latency); implement automated re-evaluation on any change (model/agent/prompt/data).Safety & Guardrails: Specify policy-as-code, entitlement checks (role/row/column), PII/PHI handling, and content moderation; define red-team tests and jailbreak defenses.Data & Interfaces: Define schemas for AI inputs/outputs; guide ontology/taxonomy alignment; ensure provenance/lineage for KG/RAG pipelines; minimize data movement.Operability & Observability: Standardize tracing/logging/metrics for AI runs (call graphs, token/latency/cost); set SLOs and error budgets; partner with DevOps for CI/CD and environment promotion gates.Technical Leadership: Review designs/PRs; mentor AI engineers; communicate decisions/trade-offs to stakeholders; maintain decision records and roadmaps.Required SkillsApplied LLM Systems: 1+ years in ML/AI or platform architecture with production LLM solutions (planning/reasoning, tool use, function-calling, agent ecosystems).Knowledge & Retrieval: Hands-on with Knowledge Graphs (RDF/SPARQL or property graph/Gremlin) and RAG (chunking, embeddings, retrieval/reranking); practical trade-offs between the two.Model & Vector Ecosystem: Experience with at least one major model platform (Azure OpenAI, Vertex AI, Anthropic, open-weights) and vector DBs (pgvector, Pinecone, Weaviate, Milvus) plus search (OpenSearch/Elasticsearch).Evaluation Engineering: Ability to construct evaluation harnesses, design rubrics, and automate regression testing; familiarity with A/B testing and human-in-the-loop review.Security-by-Design: SSO/OIDC, secrets management, least-privilege design, policy-as-code, data minimization, and auditability for AI systems.Software & APIs: Strong API design (REST/gRPC), JSON schema contracts, error taxonomies, retries/backoff/idempotency; proficiency in one or more languages (Python, TypeScript/Node.js, Go, or Java).Observability & Reliability: OpenTelemetry or equivalent for traces/metrics/logs; resiliency patterns (circuit breakers, bulkheads, backpressure); performance tuning and cost governance.Good to Have SkillsOntology & Graph Practices: SHACL, OWL, ontology stewardship, and data quality checks; graph query optimization and subgraph extraction patterns.Prompt & Tooling Ops: Prompt versioning, prompt injection defenses, retrieval parameter tuning, and structured tool-use schemas (e.g., MCP-style adapters).MLOps & Platforms: IaC (Terraform), CI/CD for models/prompts/configs, feature flags, canary releases; experience with GPU/accelerator considerations.External Tools: Designing safe, auditable action patterns and HITL approvals for external tool execution (simulation/models, document generation, ticketing).Cost/Performance Analytics: Token accounting, cache strategies, and per-agent cost ceilings; dashboards for cost-per-answer and latency P50/P95 targets.UX for Explainability: Collaborating on rationale/explanation UX so users understand sources, subgraphs, and model/tool decisions.