Sr Principal AI Architect

Job Role: Sr Principal AI Architect

Location: Bay area, CA (Hybrid)

Client  : WinWire

Salary  : $250-$300K PA + Benefits

Job Type: FTE

Key skills: Strong GenAI, LLM Production exp, RAG, Azure, (working directly with customers or business stakeholders) This is FDE role, (Forword Deployment engineer role)

Experience: 18+ Years

·                 Will work on the intelligence layer for multiple programs — owns all model quality, RAG accuracy, prompt engineering, and AI safety across applications

·                 Socratic tutor persona, adaptive learning recommendation engine, multi-modal AI (text and voice), RAG evaluation framework, and feedback loop into retrieval

·                 6-LLM call chain orchestration (NeMoGuardrails → intent classification → query rewriting → RAG → synthesis), , and compatibility check logic

·                 Production-grade AI quality from launch — this is not a research or prototyping role; accuracy thresholds, latency requirements, and safety guardrails must pass InfoSec adversarial testing before Release 1

Required Skills

Experience

·                 Total IT 18+ Years

·                 4–7 years of software engineering with at least 2 years focused on LLM application development in production — not research, not demos, not internal tools with 10 users

·                 Has shipped an LLM-powered feature or product to production where real users depend on the accuracy and the engineer owns the quality metrics

·                 Has owned an AI safety or guardrails implementation for a customer-facing product — not just added an off-the-shelf filter; designed and tested the safety layer

·                 Has built RAG evaluation pipelines and used them to make go/no-go release decisions — accuracy gating is part of the workflow.

·                 Has profiled and optimized a multi-step LLM call chain for latency

LLM Application Development

·                 LLM prompt engineering — system prompts, few-shot examples, chain-of-thought, instruction following · Expert · Must-have

·                 Multi-step LLM chain orchestration — LangChain, LlamaIndex, or custom orchestration · Expert · Must-have

·                 Multi-turn conversation design — context window management, conversation summarization, session memory · Advanced · Must-have

·                 Streaming LLM response handling — token-by-token streaming, partial response rendering · Advanced · Must-have

·                 Model selection and benchmarking — matching model size to task; balancing latency, cost, and accuracy · Advanced · Must-have

RAG Pipeline Design & Quality

·                 RAG pipeline design — chunking strategy, embedding model selection, retrieval configuration · Expert · Must-have

·                 Vector similarity search tuning — index parameters, similarity thresholds, retrieval depth · Advanced · Must-have

·                 Reranking — cross-encoder rerankers, relevance scoring · Advanced · Must-have

·                 RAG evaluation frameworks — RAGAS, TruLens, or equivalent; automated eval pipelines · Advanced · Must-have

·                 Hybrid search — combining dense vector retrieval with BM25 or keyword search · Proficient · Nice to have

AI Safety & Guardrails

·                 Prompt injection detection and mitigation · Advanced · Must-have

·                 Jailbreak testing and red-teaming LLM systems · Advanced · Must-have

·                 Content safety classifier integration · Advanced · Must-have

·                 Hallucination detection and mitigation strategies · Advanced · Must-have

·                 Topical control — enforcing scope boundaries on LLM responses · Advanced · Must-have

Evaluation & Production Quality

·                 Automated evaluation pipeline design — test set curation, metric selection, regression detection · Advanced · Must-have

·                 A/B evaluation methodology for prompt and model changes · Proficient · Must-have

·                 Latency profiling for LLM call chains — identifying bottlenecks across multi-step pipelines · Proficient · Must-have

·                 Feedback loop design — user signal collection, signal-to-retrieval-weight integration · Proficient · Must-have

·                 Production model monitoring — accuracy drift detection, quality degradation alerting · Proficient · Must-have

Development

·                 Python — ML/AI application development, async programming · Expert · Must-have

·                 API design for AI services — streaming endpoints, error handling, timeout management · Advanced · Must-have

·                 Embedding model operations — model selection, batch embedding, index updates · Advanced · Must-have

Nice to Have

·                 Adaptive learning systems or personalization engine experience

·                 Knowledge graph integration with RAG

·                 Multi-agent orchestration patterns

·                 ServiceNow API integration

·                 Prior experience building AI products on NVIDIA infrastructure