Job Description
We are building Cognium – an enterprise-grade platform for building, deploying, orchestrating, and governing AI agents at scale. As Principal Architect, you will own the end-to-end system design and make every foundational technical decision: from the distributed architecture (Kubernetes, Temporal, Kafka, NATS) to the AI-native patterns (RAG pipelines, LLM routing, multi-agent orchestration, guardrail frameworks). This is not a slide-deck architect role – you will write code, review every critical PR, and personally build the hardest subsystems.You will be the technical conscience of the platform. When the team debates Temporal vs. a custom state machine, Kafka vs. Pulsar, pgvector vs. Weaviate, or monolith vs. microservices – your judgment settles it, backed by hands-on prototyping, not just theory
System Architecture & Design Own the end-to-end architecture of the Cognium platform across all layers: API Gateway (Envoy), Agent Orchestration (Temporal), Agent Runtime, LLM Router, RAG Engine, Tool Gateway, Policy Engine, Observability, and Infrastructure. Design and maintain the logical architecture layers: Presentation → API Gateway → Security Pipeline → Orchestration → Runtime → LLM/RAG/Tools → Persistence → Infrastructure. Define the Control Plane vs. Data Plane separation: global metadata (CockroachDB) vs. per-region execution (PostgreSQL+Citus, Redis, pgvector).
Hands-On Technical Leadership Personally design and implement the most complex subsystems: LLM Router (smart routing, fallback chains, A/B testing), multi-agent orchestration engine (supervisor pattern, handoff protocol, shared scratchpad), and the security pipeline (prompt injection defense, guardrail framework). Write production code in Go (performance-critical services), Java/Spring Boot (business logic services), and Python (ML pipelines, RAG engine). Expected contribution: 40-50% hands-on coding in the first 12 months
AI/ML Architecture Design the RAG pipeline architecture: document processing → chunking → embedding → hybrid retrieval (pgvector + Elasticsearch BM25) → RRF fusion → re-ranking → citation building. Own the RAGAS quality framework (Faithfulness, Relevancy, Precision, Recall). Architect the LLM Router for model-agnostic operation: unified invocation interface across Anthropic, OpenAI, Google, Mistral, and self-hosted models (vLLM/TGI). Design routing rules, fallback chains, and A/B testing infrastructure ...
