Ai Tech Architect

Overview Own the end-to-end architecture of production AI systems with a strong hands-on bias. You’ll design robust, cost-efficient, and secure agentic/GenAI solutions on AWS. Part of your job will be to unblock lead developers by debugging code, optimizing performance, and guiding best practices. Expect to turn complex requirements into scalable, observable, and well-governed platforms. Responsibilities - Define target architectures for agentic systems (planning/reasoning/tool-calling), GenAI/RAG pipelines, and evaluation loops; produce clear design documents with Flow/UML/sequence diagrams and AWS deployment topologies. - Size and optimize infrastructure for cost and performance: model throughput/latency, concurrency, autoscaling policies, CPU/GPU needs, memory footprints, vector index sizing, storage/egress, and token budgets. - Lead deep-dive debugging and incident resolution: profile bottlenecks, fix defects, stabilize services; pair-program with developers to raise the engineering bar. - Establish reference implementations for multi-agent frameworks (Semantic Kernel preferred; LangGraph/AutoGen/CrewAI acceptable), tool/function schemas, validation, memory, grounding, and multi-step planning. - Architect retrieval and hybrid search systems: ingestion, chunking, embeddings, ranking, caching, freshness, and grounding; evaluate recall, precision, and hallucination risk. - Productionize on AWS using Amazon EKS, S3, SQS/SNS, and AWS Bedrock; integrate identity (Okta/IAM), secrets (AWS Secrets Manager), eventing, and observability; enforce SLIs/SLOs and error budgets. - Make systems observable: distributed tracing, metrics, and logs using OpenTelemetry and Datadog; standardize dashboards,alerts, and tool/trace replay. - Build evaluation and promotion workflows: prompt/flow tests, golden sets, offline batch runs, A/B experiments, regression suites, and rollout gates. - Design security and safety controls: threat modeling, prompt-injection defense, sandboxed tools, policy enforcement, red-team testing, PII/data governance, and audit trails. - Define platform standards: reusable SDKs, connectors, CI/CD templates, runbooks, and architecture review checklists. - Partner with product, data, and SRE teams to plan capacity, disaster recovery, multi-region posture, and upgrade paths; lead readiness reviews and post-incident RCAs. - Mentor engineers and review PRs with a focus on reliability, correctness, and maintainability. Must Have - 7–10 years in software/AI engineering with at least 4+ years building GenAI applications and 2+ years architecting production agentic systems. - Strong hands-on expertise in Python 3.11+ (typing, asyncio, packaging, profiling, pytest); able to dive into code,fix bugs, and optimize performance-critical paths. - Experience with one or more agent frameworks (Semantic Kernel, LangGraph, AutoGen, CrewAI) and function/tool calling with schema and argument validation. - Proven design of GenAI/RAG/hybrid retrieval systems using AWS Bedrock, OpenSearch Serverless, or other vector databases; grounding and retrievalevaluation experience. - Deep knowledge of AWS architecture: Amazon EKS, Bedrock, S3, SQS/SNS, RDS (SQL Server/PostgreSQL), ElastiCache (Redis), Secrets Manager, IAM/Okta, Kong API Gateway, OpenSearch Serverless, and Datadog. - Observability expertise: distributed tracing (OpenTelemetry), metrics, logs, correlation IDs, and service-level objectives; mature incident response and on-call practices. - Cost and performance engineering mindset: capacity modeling, GPU/CPU sizing, autoscaling (HPA), batching/streaming, caching, and FinOps discipline. - Security and safety fundamentals: least privilege, data isolation, policy enforcement, content moderation, jailbreak/PII defenses, and compliance awareness. - Excellent technical communication: clear diagrams, ADRs, design docs; empathetic, structured code and architecture reviews. Good to Have - Multi-agent orchestration patterns: task decomposition, coordinator-worker, human-in-the-loop, graph-based planning. - Deep expertise with vector databases and retrieval: OpenSearch Serverless, Pinecone, pgvector, Redis. - Evaluation frameworks: red teaming, automated guardrails, regression testing, rollout gates, canary deployments. - Frontend integration for agent UIs (streaming responses, tool traces), secure connector APIs, and AuthN/Z best practices. - Policy-as-code (OPA) and multi-tenant architecture (RBAC, quotas, usage metering). - Knowledge of Kong API Gateway, LaunchDarkly/Flipt for feature management, and NeMo Guardrails for runtime safety. - CI/CD exposure (build/test with GitHub Actions, deployments via Terraform/AWS IaC templates). Core Tech Stack (our core; equivalents welcome) - Python 3.11+, FastAPI, Pydantic v2, SQLAlchemy 2.X, Alembic, pytest. - Amazon EKS, AWS Bedrock, Amazon SQS/SNS, Amazon RDS (SQL Server/PostgreSQL), ElastiCache (Redis). - Amazon S3 for storage, Amazon ECR for container images, OpenSearch Serverless for vector storage. - AWS Secrets Manager, Okta IAM, NeMo Guardrails, Kong API Gateway. - OpenTelemetry + Datadog for observability and monitoring. - Custom RAG Services, Bedrock Knowledge Base, and LLM evaluation with Phoenix, Arize, and Promptfoo.