Product descriptions:
Furniture.com is the first place where shoppers can discover, compare and check out across multiple trusted furniture brands. In one experience, with one cart. As such, Furniture.com sits at the center of a vast, fragmented furniture ecosystem where millions of product records and reviews influence how consumers discover and purchase home furnishings. Inconsistent data and manual processes once made it hard to deliver reliable, intelligent shopping experiences at scale. By consolidating data and AI onto Databricks, Furniture.com operationalized generative AI across 4.5 million reviews and hundreds of thousands of products. AI-powered summaries and automated furniture collections now enhance the buying journey from engagement to conversion.
Scaling AI across 70+ retail partners exposes data gaps
Furniture.com operates as a one-stop shop for furniture and decor, merging more than 70 retail partners into a single-cart shopping experience. But as an aggregator, Furniture.com does not control the quality, structure or completeness of the data it receives. Each partner provides product feeds in different formats and varying levels of detail, making it challenging to deliver shoppers with consistent, engaging product information.
The team recognized that leveraging AI could turn messy data into actionable insights and a smarter shopping experience. But moving from concept to production required significant effort. As Furniture.com’s Associate VP of Data and AI, Yasel Garces, explained, “To adapt AI to your specific use case and deliver something meaningful to users, you need significant preprocessing and orchestration around the model. Calling an LLM alone is not enough. The real value comes from packaging it correctly within a well-designed pipeline.”
Scaling AI meant first solving an underlying challenge: fragmented, inconsistent and incomplete product data. Two high-impact product discovery problems — reviews and collections — made that quality gap especially visible.
Turning unstructured reviews into trustworthy insights
While many retailers host reviews on their own websites, that information does not always flow cleanly into syndicated product feeds. At scale, thousands of unstructured reviews per product created both a usability issue and an evaluation challenge. Furniture.com sought to build an intelligent review summarizer and tagger that could surface balanced, trustworthy insights without overwhelming shoppers.
Transforming disorganized products into curated collections
To create a more engaging shopping experience, Furniture.com introduced Collections — groups of coordinated furniture pieces designed to complement one another. A shopper who selects a sofa may want to see the matching loveseat or ottoman from the same line. However, retailers do not consistently provide structured collection metadata, making it difficult to consistently organize pieces in a way that makes sense for shoppers. Brianna King, Machine Learning Engineering Manager, described the ambiguity the team faced. “We have all the product titles, but there’s no clear way to identify which part of the title represents the collection. We had to build an enhanced proprietary system that could reliably extract the collection name and then group the other products that belong to that same collection.”
Across both use cases, data preparation was a central hurdle. Reviews had to be ingested from multiple vendors, deduplicated, cleaned and processed incrementally as new content arrived. Product titles required linguistic parsing to separate meaningful collection names from noise, along with change data capture to avoid reprocessing millions of records daily. Model accuracy and QA introduced another layer of complexity. For collections, there was no perfect ground truth dataset to measure traditional accuracy. “We have a strong in-house QA team that reviews outputs to ensure they actually make sense. But once a model is in production, we want that process to be automated. That’s why it’s so important for us to have robust alerting systems in place,” said Brianna.
The scale of their ambition made it clear that fragmented tools and disconnected workflows would not be sufficient, setting the stage for a unified data and AI foundation.
Powering intelligent product discovery with AI on Databricks
To operationalize AI across millions of data points and dozens of partners, Furniture.com consolidated its entire data and machine learning infrastructure on the Databricks Data Intelligence Platform. By bringing data engineering, model development, governance and serving into one platform, the team reduced architectural friction and created a production-ready AI foundation.
AI-powered review summaries and tags
As Furniture.com’s Data Engineer, Vishnu Kalakata, stated, “Furniture.com has successfully processed over 4.5M reviews across 30 partners, with coverage continuing to expand. Using Delta Lake and PySpark Databricks Jobs, we built reusable pipelines that ingest review feeds from multiple vendors, de-duplicate records, apply cleaning and transformations and implement change data capture so only new or updated reviews are processed. Through this architecture, we transformed unstructured data into a scalable, revenue-driving system.” All workflows are orchestrated through Lakeflow Jobs, allowing the team to trigger pipelines as new feeds arrive and monitor job execution in production.
Using Model Serving, the models are deployed as PyFunc models and registered in Unity Catalog, where they are versioned, tagged and promoted across environments. The first model generates AI summaries and a balanced overview of positive and negative themes. The second model analyzes sentiment and extracts structured tags that power filters on the product detail page. According to Brianna, “As soon as we deploy a model and log it to Unity Catalog, we can test changes quickly in development and promote updates without disrupting downstream workflows. With Databricks Asset Bundles and GitHub Actions, the entire CI/CD process is automated.”
By operationalizing review ingestion, summarization and tagging within a unified Databricks environment, Furniture.com transformed raw, unstructured feedback into a scalable trust engine. What was once fragmented review data is now a structured, revenue-driving asset that improves product discovery and increases engagement across partners.
Helping shoppers discover complete rooms, not just individual products
Furniture.com deployed eight models to identify and validate related furniture pieces. Natural language preprocessing and business-logic filtering are executed within Databricks pipelines, with a final validation step that leverages an LLM to confirm collections. As Yasel explained, “We architect everything in Databricks to be modular. The models exist as independent components within the overall pipeline, which means we can unplug one model and plug in another without reworking the entire system.”
The team uses MLflow to log model experiments, track versions and manage promotions between environments. During development, generative models are evaluated using MLflow’s LLM-as-judge. “We maintain dashboards that track key output metrics, including the ratio of positive to negative sentiment and any model errors. QA actively monitors these signals via Power BI dashboards to ensure the models are performing as expected,” explained Anwaar Msehli, Senior Applied ML Engineer.
By transforming ambiguous product titles into validated collections, shoppers can now easily explore complete product lines, driving higher engagement and conversion.
Building a governed foundation for continuous innovation
As AI moved into production, governance became essential. Models are versioned in Unity Catalog, experiments tracked in MLflow and pipelines are orchestrated through Lakeflow, creating end-to-end traceability across the AI lifecycle.
Unity Catalog provides centralized data lineage and access controls across both datasets and models, ensuring experimentation, deployment and monitoring occur within a single controlled environment. “Having your entire data ecosystem in one place is a game changer,” said Yasel. “Data supports AI, and AI helps improve the data. It becomes a two-way relationship that continuously strengthens the system.”
With governance embedded across data and models, Furniture.com can scale quickly without sacrificing reliability. New features and partners integrate seamlessly, and AI evolves as a continuously improving system that strengthens both data quality and shopper experience.
Doubling conversion with generative AI
With the Databricks Platform, Furniture.com operationalized generative AI across millions of reviews and enriched more than 265,000 products directly within the shopping experience.
Revenue and shopper behavior
The way shoppers interact with product detail pages has changed as AI delivers a more personalized and engaging experience. “We saw that 95% of users engaged with these intelligence features,” explained Danica Chan, Furniture.com’s Product Manager. Across products where shoppers engage with AI summaries and tags, Furniture.com observed a 2x increase in conversion rate. The summaries serve as a leading indicator, with the vast majority of shoppers reading them before making a purchasing decision.
Merchandising impact
By structuring more than 8,200 collections (12% of the catalog) into coordinated product lines, conversion increased by 26%. What began as a data enrichment effort has become a core merchandising driver, making it easier for customers to discover complementary products and build cohesive spaces.
Operational and productivity gains
Not only is the team delivering innovations, but they’re doing so at unprecedented speed. In a single year, they deployed eight production AI models with an AI team of only five people. “Not since Databricks came around have I experienced this level of collaboration between data engineers and data scientists. Bringing everything together has facilitated a real cultural change in how our technical teams work together,” said General Manager Daniel Russotto.
By unifying data and AI at scale, Furniture.com is transforming shopping into a seamless, intelligent experience — driving engagement, boosting conversion and setting a new standard for retail innovation.