Available for select engagements

Designing conversations
between humans and language models.

I'm Aria Vance, a prompt engineer and LLM systems designer. I turn fuzzy ideas into reliable, production-grade AI behavior — across chat, agents, and automated pipelines.

120+
Prompt Systems Shipped
38%
Avg. Accuracy Lift
6
Years in NLP / LLMs
14
Industries Served
About

I speak fluent model.

I specialize in prompt architecture, evaluation design, and agentic workflows for teams building serious products on top of large language models. My work sits at the intersection of linguistics, systems thinking, and product judgment.

Before prompt engineering had a name, I was writing rule-based dialogue trees and fine-tuning small transformers. Today I help teams get more out of frontier models — GPT, Claude, Gemini, Llama — without a single line of retraining.

I care most about reliability under distribution shift: prompts that don't just demo well, but hold up against messy real-world input, adversarial users, and edge cases nobody thought to test.

Focus areas
Prompt architecture · Eval design · Agent orchestration
Model families
Claude, GPT-4/5, Gemini, Llama, Mistral
Tooling
LangChain, DSPy, PromptLayer, Weights & Biases
Based in
Remote — UTC-5
Currently
Leading prompt strategy for a Series B AI startup
Capabilities

What I actually do

Prompt engineering is a systems discipline. Here's the toolkit I bring to every engagement.

Prompt Architecture

Structured, versioned prompt systems with clear roles, constraints, and few-shot scaffolding built for maintainability.

Evaluation & Testing

Golden datasets, LLM-as-judge rubrics, and regression suites that catch quality drift before your users do.

Agentic Workflows

Multi-step tool-using agents with guardrails, retries, and graceful degradation for production reliability.

Fine-tuning & RAG

Knowing when a better prompt beats a fine-tune — and building the retrieval layer when it doesn't.

Cost & Latency Tuning

Token-budget-aware prompt compression and model routing that cuts spend without cutting quality.

Red-teaming & Safety

Adversarial testing for jailbreaks, prompt injection, and hallucination under production traffic patterns.

Selected Work

Case studies

A few systems I've designed, tuned, and shipped to production.

Customer Support

Support Copilot for a fintech unicorn

Redesigned a support agent's prompt chain to cut escalations and hallucinated policy answers, using retrieval-grounded responses and strict output schemas.

RAGFunction CallingGPT-4
-42%Escalation rate
3.1xFaster resolution
Content Generation

Brand-voice engine for a media company

Built a modular persona + style-guide prompt system that let non-technical editors generate on-brand copy across 9 product lines.

ClaudeStyle TransferTemplates
92%Editor approval rate
5hr → 20minDraft time
Agentic Automation

Autonomous research agent

Designed a multi-agent pipeline for market research: planning, web retrieval, synthesis, and self-critique loops with human checkpoints.

Multi-agentTool UseSelf-critique
67%Time saved / report
0.94Factuality score
Evaluation

Eval framework for a healthcare LLM

Built an LLM-as-judge evaluation harness with clinician-reviewed rubrics to gate every prompt change before deployment.

LLM-as-JudgeCI/CDDSPy
100%Changes gated
-58%Regression rate
Career

Experience

Lead Prompt Engineer2023 — Present
Northlight AI

Own prompt strategy and evaluation infrastructure for a suite of customer-facing AI products used by 2M+ end users.

AI Product Engineer2021 — 2023
Fable Labs

Shipped GPT-3-powered writing tools; built the company's first internal prompt-versioning and A/B testing system.

NLP Engineer2019 — 2021
Verse Analytics

Built classical NLP pipelines and early transformer fine-tunes for document classification and entity extraction.

M.S. Computational Linguistics2017 — 2019
University of Washington

Thesis on discourse coherence modeling in generative dialogue systems.

"Aria doesn't just write prompts — she designs the whole reasoning process. Our hallucination rate dropped by half in one sprint."

Devon Okafor — Head of AI, Northlight AI

Let's build something the model actually understands.

Open to consulting engagements, fractional prompt-engineering roles, and interesting problems.

Email me →