What a Summer of Real Data Taught Me at Phobio

Why Is Predicting Used-Device Resale Value Important?

Used-device resale value is a critical input in determining how much a trade-in or buyback provider can offer for a device while maintaining a sustainable margin.

At Phobio, the fundamental question behind my summer project was straightforward: What will a used device resell for?

Phobio acquires used devices and ultimately moves them into the secondary market. The difference between acquisition cost and resale value means accurate pricing matters on both sides of the transaction.

Price a device too low, and you risk losing the trade. Price it too high, and you risk reducing or eliminating the margin when the device is resold.

My work focused on giving the team sharper tools for decisions they already make every day:

• A clearer view of current secondary-market pricing
• A data-driven approach to predicting resale value
• Better visibility into how devices retain or lose value over time

The goal was to help strengthen the data foundation behind device pricing decisions.

What Is Secondary-Market Intelligence?

Secondary-market intelligence is data that helps businesses understand current pricing, demand, and value trends for previously owned products. For used electronics, this can include what comparable devices are selling for or being valued at across the broader market.

My first project looked outward.

Phobio's internal data can tell us what a device sold for historically. But historical internal data alone doesn't necessarily tell us what the broader market is offering for that same device today.

That context matters when the objective is to keep device pricing competitive while protecting margin.

I built an internal market intelligence capability that gives the pricing team more timely visibility into the secondary device market. Instead of depending entirely on outside sources, the team now has a scalable foundation for gathering market pricing intelligence on its own schedule.

It was also my first major lesson of the summer: real-world pricing data is messy.

Before I even began working deeply with Phobio's own production data, I had to learn how to organize, interpret, and evaluate market data that wasn't necessarily designed for machine learning.

How Can Machine Learning Help Predict Device Resale Value?

Machine learning can help predict used-device resale value by identifying relationships within historical pricing and device data and applying those patterns to future valuation decisions.

Building the model, however, was only one part of the challenge.

My primary project began with a large production database that wasn't organized for machine learning. There wasn't a pre-built pipeline or prepared dataset waiting for me. I had to engineer the path from raw production data to a usable prediction tool.

That meant:

• Understanding a large and complex production dataset
• Identifying the information relevant to resale value
• Cleaning and structuring data for modeling
• Engineering a reliable data pipeline
• Selecting and validating model features
• Testing predictions against historical outcomes
• Turning the output into something business users could actually use

Working within a live production environment added another layer of responsibility. The system supports real business operations, so protecting data integrity and system stability was just as important as extracting useful insights.

By the end of that process, I had transformed a sprawling data source into a structured foundation for resale value modeling.

Why Does Feature Selection Matter in Resale Value Prediction?

Feature selection determines which pieces of information a machine learning model uses to make a prediction. Choosing relevant, reliable features is essential because a model can appear accurate during testing while relying on relationships that won't hold up in real-world conditions.

That became one of the most important lessons of my internship.

The technical build was only part of the work. The more subtle challenge was learning what the data could actually tell me and recognizing when a feature that looked promising shouldn't be trusted.

I spent significant time testing potential inputs and rejecting those that failed to remain reliable under real operating conditions.

The objective wasn't simply to create a model that performed well in isolation. It was to develop something that could remain useful as market conditions changed.

How Does Original Retail Price Help Measure Device Depreciation?

Original retail price provides a baseline for measuring how much value a device retains or loses over its lifecycle. Comparing launch price with current resale value can help quantify device depreciation and residual value over time.

During the project, I discovered that our production environment didn't contain a complete reference for original device retail prices.

That was an important missing piece.

To address it, I built a comprehensive launch-price reference dataset covering nearly every device family relevant to the project.

I also assigned confidence scores to individual entries so future users of the dataset could understand the reliability and limitations of the underlying information.

What began as a feature required for my resale value model became something broader: a standalone internal data asset that can support analysis beyond the original project.

What Did the Data Science Project Deliver?

By the end of the summer, the individual pieces had come together into a broader internal pricing capability.

The project included:

• Secondary-market intelligence tools to provide greater visibility into current market pricing
• A resale value prediction model validated against a large historical dataset
• Supporting reference datasets to strengthen device-level analysis
• Data pipelines that transformed production information into usable modeling inputs
• An internal application that made the resulting insights accessible to business users

I closed out my internship by presenting the work to Phobio's leadership team.

Instead of starting with algorithms or model architecture, I framed the presentation around the business problem: pricing, margin, and cost.

Learning to translate technical work into business impact became another important part of the experience.

What Does It Take to Build a Useful Machine Learning Model?

I came into the internship thinking the model would be the main event.

It wasn't.

The model itself represented only a fraction of the work required to create something useful. The real value came from everything surrounding it:

• Understanding the business problem
• Studying the secondary market
• Mapping a complex production data environment
• Cleaning and structuring real-world data
• Engineering reliable data pipelines
• Selecting defensible features
• Building missing reference datasets
• Validating model performance
• Making the output accessible to end users

A machine learning model is only as useful as the data, infrastructure, context, and decision-making process surrounding it.

That's probably the biggest technical lesson I'll take away from the summer.

What I Learned From the Phobio Team

The technical work was important, but the people are what I'll remember most.

From my first day, I had the opportunity to work directly with Eric Attanasio, Phobio's Chief Product Officer, and Winston Astrachan, Phobio's Chief Technology Officer. Both were accessible throughout the summer, and conversations with Eric helped me understand not only the project but the operational realities behind the data.

I never felt like I was being handed an isolated intern project. I was working on a real business problem with people who understood how the results could be used.

That experience extended across Phobio.

Missy Taylor, Phobio's CEO, made time for me, and I had the opportunity to meet people across operations, accounting, product, technology, and other parts of the company. Every conversation gave me another piece of the picture of how the business works.

The kindness of Denny Juge and Amber Butler stood out from the moment I met them. Even lunches with the team became part of the experience.

Phobio didn't make me feel like a temporary worker. I felt like part of the team.

What I Took From My Data Science Internship

I came to Phobio knowing how to train a machine learning model.

I left understanding what it takes to build the pipeline, datasets, market context, validation process, and product around one.

More importantly, I learned that working with real business data requires judgment just as much as technical ability. You have to understand what the data means, recognize what it can't tell you, and build systems that account for those limitations.

I'm grateful Phobio gave me a problem with real stakes and the room to own it from beginning to end.

The experience changed how I think about data science. The goal isn't simply to build a model. It's to build something people can trust and use to make better decisions.