Multi-agent system design : supervisor–worker hierarchies, planner–executor and reflection loops, blackboard and swarm patterns; task decomposition, delegation, and inter-agent communication protocols; deciding when a single-agent vs. multi-agent topology is architecturally justified.
Agent state, memory & context engineering : short-term vs. episodic vs. semantic memory design, checkpointing and resumability, durable execution for long-running agents; context-window budgeting, compaction/summarization strategies, and retrieval-augmented context assembly.
Framework and protocol depth : LangGraph (graph state machines, interrupts, human-in-the-loop nodes), CrewAI, AutoGen/Semantic Kernel; MCP (Model Context Protocol) for tool and resource federation and A2A for agent interoperability; sound judgment on custom orchestration vs. framework adoption.
Model strategy & token economics : model portfolio design and routing (frontier LLMs vs. SLMs), structured outputs and function-calling schema design, constrained decoding; fine-tuning vs. RAG vs. prompt-optimization trade-offs; prompt caching, batching, distillation, and quantization to hit latency and cost SLOs.
Retrieval & knowledge architecture : hybrid retrieval (sparse + dense), rerankers, GraphRAG and knowledge graphs; chunking and embedding strategy, freshness pipelines, and access-control-aware retrieval (document/row-level security) for regulated enterprises.
Evaluation architecture : golden datasets, LLM-as-judge with calibration, trajectory-level agent evals, regression harnesses wired into CI/CD gates, and online canary/A-B evaluation for continuous quality assurance.
Guardrails, safety & governance : prompt-injection and jailbreak defenses, PII detection/redaction, policy engines, sandboxed tool execution, human-approval gates for high-risk actions, and full audit trails/lineage for responsible AI and regulatory compliance.
Production & platform architecture : model gateways, multi-tenancy, VPC/private endpoints, HA/DR, autoscaling, rate limiting, and circuit breakers; observability via distributed tracing (OpenTelemetry), token/cost telemetry, and drift monitoring at enterprise scale.
Enterprise integration : event-driven and API-led integration patterns, identity propagation (OAuth/OIDC), secrets management, and integrating agents with CRM, contact center, workflow platforms, and legacy estates.
Multimodal & emerging stacks : voice agents (streaming ASR/TTS – e.g., ElevenLabs), avatar/video (HeyGen), computer-use agents; fluency with AI-native tooling (Claude Code, Cursor) and evolving OpenAI/Anthropic platform capabilities.
