Challenge

The first challenge was cleaning the raw dataset. It was real appraisal data, which meant it was messy, inconsistent, and chaotic. There are a few different paths you can take to parse this kind of data such as LLMs or regexes. I went the regex route, which meant more manual work, but it gave me tight control over the process with faster parsing time at zero cost. I still used an LLM to help extract structure from the “public remarks” text fields, filling in missing values where possible to improve data quality.

Once the data was cleaned, I moved on to feature engineering. The core idea here was to calculate differences between the subject property and each candidate comp for features like gross living area, lot size, number of bedroomss and bathrooms, etc.. While the initial implementation was relatively straightforward, this component was also that involved the most creativity. Do you keep it simple? Do you include more abstract metrics? What adds signal vs. noise? I started with basic differences to keep the system modular, then gradually layered in more advanced features and used SHAP analysis to evaluate which ones mattered most.

For the model, I used an XGBoost ranking model. It takes a subject property and a set of candidates and outputs a relevance score for each one. It was a natural fit for this type of pairwise ranking task. It's accurate, fast, and easy to iterate on.

With the model selecting top comps, the next step was generating explanations. Real appraisers typically select 3 comps and justify their choices with written reasoning. To mirror this, I took the top 3 model-ranked comps, bundled their key feature diffs and SHAP values, and sent that context to a GPT-3.5 instance tasked with generating natural-language explanations. The goal was provide clear, human-readable justifications for each selection.

At this point, I had a working ETL pipeline, a trained model, and an explanation generator. The final step was building a feedback loop. Appraisers can review the model's selections, approve or reject individual comps, and submit feedback. Approved comps are used to retrain the model, improving it over time. There's also a reset option to restore the system to its baseline state if needed.

Surprisingly, the one of the hardest parts of the project wasn't the model or the data, it was designing a frontend that made all of this feel intuitive. I ended up building a clean, lightweight Streamlit interface that shows all the relevant data like subject and comp features, price range recommendations, and buttons for feedback and reset.The result is a tool that feels simple but really is an intelligent appraisal system under the hood.