Principal AI Architect, Agentic Platform

About the Role We are looking for a Principal AI Architect, Agentic Platform to define the technical foundation for 6sense's next generation of agentic AI. This is the most senior AI individual contributor role in Engineering. You will establish the architecture, design patterns, runtime standards, evaluation practices, security controls, and governance mechanisms that enable teams across 6sense to build and operate production-grade AI agents safely and efficiently. You will own the architectural direction for our agentic platform, including orchestration, tools and skills, memory and context, evaluation and observability, security and privacy, guardrails, AI governance, and model strategy. This is a hands-on Principal role. You will not manage a team, but you will influence architecture across the organization through reference implementations, RFCs, technical reviews, prototypes, mentorship, and deep involvement in complex production systems. What You'll Own Agentic Architecture & Design Patterns Define the platform-wide architecture for how agents are composed, orchestrated, and executed. Establish clear boundaries between reasoning, orchestration, state, tools, and execution. Create and maintain an agentic design-pattern catalog covering approaches such as ReAct, plan-and-execute, orchestrator-worker, routing, prompt chaining, fan-out/fan-in, reflection, evaluator-optimizer, and multi-agent handoffs. Guide teams toward the simplest architecture that solves the problem, introducing multi-agent systems only when there is a clear technical need. Design memory and context architectures covering working, episodic, and semantic memory; context compaction; retrieval grounding; subagent isolation; and tenant-scoped context. Publish reference architectures, RFCs, implementation guides, and reusable patterns. Lead architecture reviews for agentic systems across Engineering. Agentic Runtime & Developer Platform Own the architecture and standards for LangGraph , including typed state graphs, conditional routing, subgraphs, checkpointing, durable execution, human-in-the-loop workflows, streaming, replay, and time-travel debugging. Establish engineering standards for LangChain and determine where framework abstractions provide value versus direct model/provider SDKs. Build and evolve the LangSmith practice for tracing, evaluation, prompt management, regression testing, annotation, and experimentation. Define the platform contract for agent development, including SDKs, scaffolding, templates, CI/CD gates, replay tooling, and production rollout patterns. Establish safe deployment mechanisms such as shadow runs, canary releases, evaluation gates, and emergency kill switches. Maintain portability across model and infrastructure providers. Tools, Skills & Agent Execution Own the agent tool and skill registry as a platform capability. Define standards for typed tool contracts, versioning, discovery, naming, descriptions, permissions, deprecation, and lifecycle management. Architect the MCP layer for 6sense data and actions and establish standards for safely consuming third-party MCP services. Address tool-poisoning and untrusted-tool risks as part of the platform architecture. Design safe execution patterns using idempotency, retries, compensating transactions, rollback strategies, rate limits, permissions, and spend controls. Define agent-to-agent delegation and handoff contracts with authenticated delegation, scope narrowing, and preserved accountability. Long-Running & Complex Agentic Workflows Architect durable, resumable workflows that can survive failures, interruptions, and long execution windows. Establish patterns for checkpointing, deterministic replay, failure isolation, saga/compensation, and recovery. Define runtime governance including limits on steps, tokens, execution time, cost, and loops. Establish stop conditions and escalation mechanisms for stuck, uncertain, or unsafe agents. Design human-in-the-loop workflows based on risk tiers while keeping approval experiences practical for users. Define reliability metrics such as task success rate, tool-call precision, groundedness, citation accuracy, latency to first useful action, unrecoverable failure rate, and cost per successful outcome. Security, Privacy & Trust Own the security architecture and threat model for agentic systems. Apply frameworks such as the OWASP LLM/Agentic Top 10 and MITRE ATLAS to production agent workflows. Design protections against prompt injection, excessive agency, tool and memory poisoning, insecure output handling, and supply-chain risks. Define agent identity and authorization models using short-lived credentials, least privilege, scoped access, secrets isolation, sandboxing, and egress controls. Ensure agents can never access data beyond the permissions of the user or system initiating the workflow. Partner with Security and Infrastructure to ensure multi-tenant isolation, auditability, detection, and incident response extend to agent execution. Establish adversarial testing and red-team practices, including jailbreak suites, prompt-injection corpora, leakage probes, and security regression tests. Guardrails, PII & Enterprise AI Controls Architect guardrails as enforceable, version-controlled policy rather than relying solely on prompts or procedural controls. Build input/output validation, safety classification, groundedness checks, schema enforcement, citation validation, and deterministic policy controls. Establish controls for AI-generated go-to-market content, including claim substantiation, disclosures, consent, suppression rules, and jurisdictional requirements. Define which actions can be automated and which require human approval. Architect PII detection, minimization, redaction/tokenization, field-level encryption, and vendor retention controls across the agent lifecycle. Prevent sensitive data leakage through traces, evaluation datasets, prompt caches, memory, vector indexes, and debugging systems.