You will build the shared platform layer that makes our agentic AI extensible, measurable, and self-improving. This is the foundation beneath everything: the orchestration engine, the skills registry, the evaluation harness, the guardrails system, the memory service, the integration gateway, and the tracing infrastructure - consumed by both the real-time conversational and offline video generation runtimes.
Your work is what separates it works for one customer with engineering involvement from it works for many customers across many domains without bespoke engineering per customer. You build the machinery once; customer-facing teams and eventually customers themselves extend it.
You will work across the platforms core systems: agentic orchestration (planning, sub-agents, parallel execution under deadlines), skills registry, evaluation and tracing, guardrails, integration gateway, memory service, and retrieval infrastructure. The shared machinery that both the real-time and offline runtimes consume.
Requirements: 4+ years building production backend/platform systems - distributed services, APIs, async processing, and systems that serve multiple consumers. Python primary; experience with high-throughput, low-latency services.
Deep hands-on experience with LLMs in production - not just using them, but building the infrastructure around them: orchestration, tool calling, retrieval pipelines, context management, prompt chaining, and multi-agent coordination. Proficiency with LangChain, LangGraph, or equivalent orchestration frameworks.
Experience building platform primitives - registries, gateways, evaluation frameworks, tracing systems, or similar shared infrastructure that other teams build on top of. You understand contracts, versioning, and what it means to ship a platform rather than a feature.
Demonstrated ability to make build vs. adopt decisions - you have integrated open-source tooling into production systems, understood its boundaries, and built the proprietary layer where needed.
Experience with RAG systems at depth - indexing strategies, retrieval evaluation, chunking, re-ranking, hybrid search, and the failure modes (context pollution, retrieval misses, conflicting sources). Familiarity with vector databases (Pinecone, Weaviate, Qdrant, or similar) and embedding models.
Familiarity with real-time system constraints - you understand how latency budgets, deadlines, and streaming shape architecture differently than batch systems.
Strong systems design skills - you can design a registry, a gateway, a memory service, or an evaluation harness that serves multiple consumers with different needs (internal teams and external customers, real-time and offline paths).
Nice to have:
Experience with evaluation of LLM systems - designing metrics, building test harnesses, measuring behavioral consistency, detecting regressions. Familiarity with DeepEval, Ragas, or similar frameworks.
Experience with multi-tenant platforms - tenant isolation, scoped configuration, and the engineering discipline required when one customers data/behavior must never leak to another.
Background in speech/multimodal AI pipelines - ASR, TTS, avatar rendering, or video generation pipelines. Understanding how these compose with a reasoning layer.
Experience with observability and tracing at the platform level - structured per-decision traces, not just application logs. Familiarity with Langfuse, Opik, OpenTelemetry, or similar.
Familiarity with agentic AI frameworks and patterns - CrewAI, AutoGen, Semantic Kernel, MCP protocol, and the failure modes unique to multi-step AI systems (planning loops, tool selection, delegation, parallel execution).
Experience with guardrails and safety for AI systems - input filtering, output validation, grounding checks, and the latency trade-offs of validation in real-time streaming.
This position is open to all candidates.