Principal AI Engineer

Principal AI Engineer

BuzzBoard builds AI products for the B2SMB market — helping agencies, media companies, and sellers understand small businesses and market to them at a level of personalization that wasn't previously economical. Our platform is built on frontier models end to end. Not AI features bolted onto a legacy product — the products are agent systems. Our products span multi-agent marketing content generation, real-time AI voice intake, and pre- and post-sales intelligence for SMBs — Zylo, IRIS, Ignite, and Ember. They share a common internal pipeline for orchestration, retrieval, tool use, evaluation, and deployment. This role owns the architecture underneath that pipeline. Not one feature — the shared layer every product depends on: how retrieval is built and measured, how agents reason and hold state and fail safely, which models run where and what happens when one degrades, how output quality is evaluated before it ships, and where the line sits between what a model decides and what deterministic code decides. You'll set that architecture, build the reusable patterns products inherit, guide the engineers implementing them, and own whether it holds up in production at volume. It's a senior individual-contributor role. Two boundaries, stated plainly:

• Not a research role. The deliverable is a shipped, reliable capability, not a paper or a demo.
• Not a DevOps role. You need to be fluent in deployment and able to get a service into production, but full-time platform ownership sits elsewhere. Your center of gravity is AI system design.

If your GenAI experience is mostly notebooks and proofs of concept that never carried real traffic, this is not the role. What you'll own AI system architecture. Design the systems behind content generation, business intelligence, recommendations, and agentic workflows — including where the boundary sits between what a model decides and what deterministic code decides. Create reusable patterns, prompts, evaluation flows, and orchestration layers that hold up across products rather than being rebuilt per feature. Agentic workflows. Lead the design of agents that reason, call tools and APIs, manage state, checkpoint, and hand tasks off cleanly. Build the guardrails that keep them grounded and predictable — the hard part is not making an agent act, it's making it fail safely and legibly when it's wrong. Retrieval and model strategy. Design RAG pipelines end to end: chunking, embeddings, metadata, reranking, and retrieval evaluation that actually measures whether the right context arrived. Own model selection across ecosystems on the axes that matter — cost, latency, accuracy, reliability — and define the fallback and model-switching behavior for when a provider degrades. Evaluation and quality. Stand up the evaluation frameworks that decide whether an AI output ships: output quality, edit ratio, hallucination rate, schema adherence, latency, failure rate, inference cost. Build regression testing so a prompt or model change can't silently break production. Turn "the output feels off" into a measured threshold. Production readiness. Partner with engineering and platform teams so systems are deployable, observable, and maintainable. Package services (Python, FastAPI/Flask, Docker) when it's the fastest path, and diagnose the production failure modes specific to AI — rate limits, cost spikes, model failures, degraded output — rather than escalating them blind. Technical leadership. Mentor GenAI engineers, review designs and prompts and workflows and evals, and translate product requirements into architecture with measurable acceptance criteria. Communicate the tradeoffs to engineers and to leadership in language each can act on.

Required

• 5+ years in engineering, AI, ML, or data-product work; 3+ years hands-on with GenAI / LLM systems
• Production GenAI experience — non-negotiable. You have built or scaled systems that carried real traffic, and you can walk through one in detail: what broke, what it cost, what you changed.
• Deep working knowledge of LLMs and SLMs: prompt engineering, structured outputs, tool and function calling
• Hands-on with at least two major LLM ecosystems, and able to reason about the tradeoffs between them rather than defaulting to the one you know
• RAG built for real: vector stores (Chroma, Pinecone, Weaviate, FAISS, or equivalent), embeddings, semantic search, and retrieval quality you've actually measured
• Hands-on with at least one agentic framework (LangGraph, CrewAI, AutoGen, Semantic Kernel, or equivalent) and the underlying concepts — memory, state, tool integration, failure handling — well enough to move across frameworks
• Strong Python; comfortable with REST APIs, Docker, a cloud platform, and basic CI/CD
• Demonstrated evaluation work: regression testing, hallucination checks, schema validation, output scoring — you can define metrics for quality, reliability, cost, and business impact
• Track record guiding a small team or owning AI architecture end to end
• Comfort creating structure in a fast-moving environment with shifting requirements

Preferred

• Fine-tuning or supervised training workflows
• SLM and open-source model deployment; model serving with vLLM, Ollama, or TensorRT-LLM
• Multimodal work across text, image, audio, or video
• Kubernetes or serverless deployment
• Evaluation/observability tooling — LangSmith, MLflow, Weights & Biases, or equivalent
• Marketing technology, SMB intelligence, or content-automation domain experience
• Responsible AI, privacy, security, and compliance practice
• Experience scaling AI systems that generate high volumes of content, recommendations, or insights

What we care most about

• You've built or scaled real GenAI systems, not just demos
• You design AI workflows that are reliable, testable, and cost-aware from the start
• You make practical model, prompt, retrieval, and architecture calls and can defend them
• You know how to balance model intelligence against deterministic logic and engineering guardrails
• You stay hands-on while guiding engineers — neither purely a manager nor purely an IC
• You think past "which model" to the whole system around the model

What this role is not

• Not a research or publications role
• Not full-time DevOps or platform ownership
• Not a role for someone whose GenAI track record is demos that never took production traffic

What we offer

• Fully remote
• A production GenAI foundation already carrying real volume — you scale it, you don't start from zero
• Genuine architectural ownership over the agentic systems that come next
• A small, high-context team that moves quickly and argues about the work
• Real impact on small businesses

Apply if you've shipped GenAI systems that carried real traffic and you want to own the architecture of what comes next.