About the Role
Context / Prompt EngineerLocation: United States (Client Onsite Hybrid (Mountain View, CA)/ Remote)
About FractalFractal Analytics is a strategic AI partner to Fortune 500 companies with a vision to power every human decision in the enterprise. Fractal is building a world where individual choices, freedom, and diversity are the greatest assets. An ecosystem where human imagination is at the heart of every decision. Where no possibility is written off, only challenged to get better. We believe that a true Fractalite is the one who empowers imagination with intelligence. And that it will be such Fractalites that will continue to build the company for the next 100 years. Fractal has been featured as a Great Place to Work by The Economic Times in partnership with the Great Place to Work® Institute and recognized as a 'Cool Vendor' and a 'Vendor to Watch' by Gartner.
Please visit Fractal | Intelligence for Imagination for more information about Fractal.
Role OverviewFractal is seeking an experienced Context / Prompt Engineer to architect, optimize, and govern dynamic context pipelines, system instructions, semantic grounding layers, and multi-agent prompt workflows for a leading global technology client across enterprise cloud, foundation model, and agentic AI initiatives.
As generative AI systems evolve from single-turn chatbots into complex, stateful multi-agent architectures, prompt engineering has expanded into Context Engineering. Treating the LLM's context window as a compiled, stateful system with a finite attention budget. Sitting at the intersection of applied AI research, software engineering, and domain data architecture, you will design modular domain context layers (separating generic context, domain-specific schemas/glossaries, and operational instructions), optimize Retrieval-Augmented Generation (RAG) and NL2SQL pipelines over enterprise cloud data warehouses, and build automated prompt optimization loops.
You will partner closely with AI architects, data scientists, evaluation engineers, and domain SMEs to ensure that LLM-powered agents deliver high instruction-following fidelity, zero hallucination over certified semantic layers, low latency, and optimal token efficiency in production.
What You'll DoSystem Prompt Design & Meta-Prompt Calibration
• Design, author, test, and maintain production system instructions, persona/role definitions, tool-calling schemas, and structured output templates (using XML tags, JSON schemas, and Markdown hierarchies) for frontier foundation models and enterprise agent runtimes.
• Curate high-signal canonical few-shot examples, chain-of-thought (CoT) reasoning trajectories, and ambiguity-resolution patterns (e.g., clarifying thin user queries and gracefully declining out-of-scope asks).
• Develop meta-prompts for synthetic training/evaluation data generation, automated insight extraction, transcript summarization, and gap analysis workflows.
Context Engineering, Compaction & Semantic Layer Grounding
• Architect tiered context assembly pipelines using modern multi-agent development frameworks—cleanly separating durable session/memory state from scoped per-turn working context views to prevent "context rot" and attention dilution.
• Structure modular domain context repositories—separating reusable General Context, Domain-Specific Context (e.g., sales, forecasting, finance models, and product glossaries), and Task Instructions—so multi-agent orchestrators inject only the precise context required for each sub-agent.
• Implement sliding-window context compaction, stale tool-output pruning, externalized artifact handles, and immutable prefix ordering to maximize prefix caching and minimize P95 latency and token cost.
RAG, NL2SQL & Multi-Agent Orchestration
• Optimize hybrid Retrieval-Augmented Generation (RAG) and semantic context retrieval over vector databases, cloud data warehouse semantic views, and BI metric definitions to ground conversational analytics and NL2SQL compilation strictly in certified enterprise data.
• Design scoped context handoffs across multi-agent architectures (e.g., Intent Router, Planner/Orchestrator Agent, RAG Context Agent, Domain Classifier, Insight Extractor, and SQL Compiler) so sub-agents operate with clean, isolated context windows.
• Integrate safety and security guardrails (such as enterprise safety filters, automated PII/DLP redaction and re-identification, and prompt-injection defenses) directly into the context pre-processing pipeline.
Automated Prompt Optimization (APO) & Continuous Improvement
• Build programmatic prompt refinement and closed-loop optimization workflows using automated prompt optimization tools and frameworks such as DSPy, OPRO, or TextGrad.
• Analyze human-in-the-loop (HITL) SME feedback, thumbs up/down signals, and production override logs to diagnose failure modes (hallucinations, schema drift, redundant insights, or sycophancy) and stage versioned Prompt Registry pull requests.
What We're Looking For
• Systems