PRISM | Proxy Recognition and Inclusive Scoring Method

Project Overview

Fall 2025 Showcase LLM Bias Evaluation Resume Screening

PRISM investigates how large language models respond to implicit demographic cues during resume screening, and how evaluator framing (strict merit, culture fit, inclusive framing) changes model decisions.

Our framework measures context dependent bias shifts, capturing patterns that static fairness tests often overlook. We use structured prompts, controlled proxy features, and fairness metrics so that shifts in model decisions can be precisely measured and compared across realistic evaluator mindsets.

Methods

We curated a structured resume dataset embedding proxy features such as neighborhood indicators, volunteer signals, and activity descriptions that correlate with real demographic patterns without naming groups directly.

These resumes were scored under multiple LLM prompt framings that represented different evaluator mindsets including strict merit focus, culture fit emphasis, time pressured triage, and inclusive hiring framing. PRISM logs how scores and rankings change under each frame and applies fairness metrics to quantify when a frame amplifies or dampens bias.

Curated resume corpus with encoded proxy features.
Multiple scoring prompts representing real evaluator instructions.
Structured logging of scores, rationales, and ranking changes.
Fairness metrics applied across proxy groups and frames.

Key Findings

Resume scores were tightly grouped overall, but a small and consistent pattern emerged in name based proxies.
Candidates with “White coded” names scored slightly higher on average than those with “Black coded” names.
Prompting context changed fairness outcomes: some neutral framings reduced disparities, but did not fully eliminate them.
Models that look fair under a single static evaluation can reveal hidden disparities when evaluated across multiple realistic frames.
Bottom line: small biases appear or disappear depending on prompt framing, which means audits must be scenario driven, not one dimensional.

Flash Video · Key Storyline

Our flash video introduces the intuition behind PRISM. It walks through how seemingly small prompt changes can make proxy based biases appear, disappear, or shift to a different part of the score distribution.

Showcase Talk & Slide Deck

The PRISM slide deck expands on the video and poster by laying out our experimental setup, quantitative results and fairness metrics.

Download Slide Deck (PPTX) View Poster (PNG)

About the Researchers

Crystal Tubbs

AI Solutions Architect & Emerging Technologies Specialist

I design and build human centered AI systems with a focus on safety, equity, and practical impact. My work spans applied AI research, fairness evaluation, representation analysis, and subliminal learning investigations across four leading AI lab ecosystems and academic environments. I am especially interested in how models learn and how people engage with them, and I use those insights to create systems that are both reliable and inclusive.

Before founding Metamorphic Curations, I served a period in the United States Army, an experience that shaped my adaptability, discipline, and ability to perform under pressure. I went on to run two successful businesses, including a mobile corporate wellness company and a long standing ecommerce operation. I also worked as a crypto trader and portfolio manager for high net worth individuals, gaining firsthand exposure to complex risk environments, data driven decision making, and high demand operational workflows. These experiences gave me a unique window into what business owners and fast moving teams truly need from technology and how AI and automation can remove real world friction.

At Metamorphic Curations, I build end to end AI systems and rapid prototypes serving a wide range of use cases including veteran support tools, logistics and operations workflows, franchise level salon automation, small business intelligence platforms, family law practice systems, coffee shop chain automation, and solutions for trades and service providers. I specialize in identifying high value use cases, designing tailored architectures, deploying production ready tools, and onboarding users in ways that drive adoption and measurable outcomes.

My mission is to build technology that moves us toward a more equitable world. I believe responsible, human centered AI is not only possible but profitable, and that systems designed with fairness at their core create stronger outcomes for everyone who contributes to them.

I currently work as a consultant and contractor and welcome full time opportunities where I can contribute to research, system design, responsible AI strategy, or emerging technology innovation. I’m open to remote roles, hybrid work in the Tampa Bay area, up to 30 percent travel, and relocation for the right long term fit.

Email LinkedIn GitHub

Destiny Raburnel

Data Engineer

Destiny Raburnel is a data engineer with a multidisciplinary foundation in biomedical engineering, software engineering, and artificial intelligence. She works across data engineering, applied machine learning, automation, and software development, building systems that combine analytical rigor with creative engineering. Her experience includes IT finance engineering, software architecture, embedded systems, academic research, AI consulting, and hands on prototyping, giving her a wide ranging perspective on how data and technology connect across an entire ecosystem.

She holds a BS in Biomedical Engineering and an MS in Software Engineering and is currently completing an MS in Artificial Intelligence, where her work focuses on intelligent systems, NLP, applied machine learning, and modern AI integration. In her current role, she designs and maintains robust data pipelines, builds APIs, develops automation workflows, and supports complex backend services. She works extensively with JavaScript, TypeScript, Python, Java, SQL, cloud tooling, and modern data engineering platforms, using them to create scalable and reliable systems that support real world decision making.

Destiny enjoys exploring emerging technologies and often experiments with robotics concepts, NFC integrations, and hands on engineering projects. She spends a significant amount of time 3D printing and prototyping because it allows her to blend creativity with technical design, and it keeps her connected to the physical side of engineering. Her interests span machine learning, embedded logic, automation, cybersecurity, and fintech, and she actively explores how these fields interact and shape the future of technology.

She currently works as a data engineer and deeply values the work she does, but she is also open to opportunities that support continued growth, innovation, and professional development. She is passionate about creating technology that is intelligent, adaptable, and user centered, and she approaches each project with curiosity, precision, and a strong sense of craftsmanship.

Email LinkedIn GitHub

Repository & Assets

The PRISM repository will include our evaluation templates, anonymized resume schema, analysis notebooks, and plotting scripts so that other teams can adapt the framework to their own models and hiring domains.

GitHub Repository