BigQuery AI Hackathon: Celebrating Innovation and a Look at What's New

The BigQuery AI Hackathon, a six-week event that concluded on September 22, 2025, brought together thousands of developers from across the globe to build innovative solutions using BigQuery's cutting-edge AI capabilities. With 5,350 entrants and 277 submissions, the hackathon was a phenomenal success, showcasing the power and versatility of BigQuery AI.

Participants were challenged to tackle real-world business problems using BigQuery’s generative AI, vector search, and multimodal features. The results were nothing short of inspiring, with developers creating everything from hyper-personalized marketing engines to intelligent triage bots.

While the competition is over, the platform is evolving fast. Below, we’re highlighting the winning projects that set the bar, followed by a breakdown of the latest SQL-based AI features, including a massive performance update, that you can use to build similar solutions today. Ready to build your own innovative solution? Get started with BigQuery AI on Google Cloud.

And the Winners Are...

We’re thrilled to announce the winners of the BigQuery AI Hackathon! These projects stood out for their creativity, technical excellence, and impactful solutions.

The AI Architect: Precision oncology at scale

This category focused on building intelligent business applications, and the winner  impressed the judges with an innovative approach to solving a complex problem.

OncOmix AI aids medical professionals by unifying patient data and medical literature within BigQuery. Using BigFrames, the system orchestrates a RAG pipeline that matches a patient’s unique genetic profile to relevant scientific papers. GeminiTextGenerator then synthesizes this data into summaries with citations, highlighting the scientific rationale for potential treatment paths.

• Watch the interview: The AI Architect

The Semantic Detective: AI-powered speech therapy

The Semantic Detective category challenged participants to uncover deep, semantic relationships in their data. This winning project demonstrated a masterful use of BigQuery's vector search capabilities.

SpeakAura AI processes unstructured audio data to analyze stammering patterns, utilizing vector search for semantic pattern matching to find similar cases from historical speech data. It generates embeddings from speech transcripts with ML.GENERATE_EMBEDDING and then uses the VECTOR_SEARCH function to perform a semantic search to find the top-k similar rows from the stored analysis results. This allows for deep speech pattern matching.

• Watch the interview: The Semantic Detective

The Multimodal Pioneer: Uncovering connections in digital services

In the Multimodal Pioneer category, participants were tasked with combining structured and unstructured data to unlock new insights. 

The winning project, TriLink, is an automated ticket handling system for home security issues that uses multimodal analysis. By creating a secure, queryable reference to customer-submitted images in Google Cloud Storage using ObjectRef, TriLink is able to use AI.GENERATE_TABLE to analyze both the unstructured text of an issue description and the submitted image. This allows the system to automatically triage issues, determining if a technician is needed on-site, and even provides DIY instructions for the customer if a site visit isn't necessary. 

• Watch the interview: The Multimodal Pioneer

What's New in BigQuery AI: Supercharging Your Solutions

The innovation doesn't stop with the hackathon. We're constantly adding new features to BigQuery AI to empower developers to build even more powerful and intelligent applications. 

For the AI Architect: Building smarter, faster applications

We recently introduced AI.IF, a powerful function that allows you to perform semantic filtering and build sophisticated business logic directly within BigQuery. And now, we're taking it a step further.

We are thrilled to announce that AI.IF now has a sublinear scaling capability (in Preview) which provides orders of magnitude improvement in performance. This new optimization dramatically reduces the number of calls to the large language model (LLM), resulting in queries that used to take an hour now taking just a couple of minutes.

You can sign up for the preview of this exciting new capability here.

In addition to AI.IF, BigQuery AI released the AI.CLASSIFY function, another powerful tool for AI Architects. This function simplifies text classification tasks, allowing you to, for example, automatically categorize support call logs with a single function call.

For the Semantic Detective: Unlocking deeper insights with less effort

The Semantic Detective category focused on leveraging vector search to find hidden relationships in data. Our new features make this process more accessible and powerful than ever:

• Autonomous Embedding Generation & AI.SEARCH: We are streamlining the entire workflow from embedding generation to search. Autonomous Embedding Generation (in preview) automates the creation of embeddings for your text data. This foundation unlocks the new AI.SEARCH function, which allows you to perform semantic searches over that data with a single function call.

• AI.SCORE: Finding data is only the first step; ranking it is the second. While AI.SEARCH (or VECTOR_SEARCH) helps you find the nearest neighbors (the "haystack"), AI.SCORE could help you zero in on the needle. You can pipeline these functions, first retrieving broadly relevant results with vector search, and then using AI.SCORE to re-rank them based on nuanced questions like "how urgent is this?" or "how actionable is this feedback?"

• AI.EMBED: For practitioners needing ad-hoc vector generation outside of an automated pipeline, the new AI.EMBED function provides a streamlined way to generate embeddings directly within your SQL queries, making it easier than ever to uncover deep semantic insights.

You can learn more in our vector search documentation.

For the Multimodal Pioneer: Bridging the gap between data types

The Multimodal Pioneer category challenged developers to work with a mix of structured and unstructured data, and the latest features can be a powerful addition to any multimodal workflow.

These new generative functions, AI.IF, AI.SCORE, and AI.CLASSIFY, can be used to analyze unstructured data stored in Object Tables. It's worth noting the distinction in input types: AI.IF, and AI.SCORE are multimodal, whereas AI.CLASSIFY is text-only input. However, you can easily bridge this gap. For instance, you could use a multimodal model to generate text descriptions of images and then use AI.CLASSIFY to categorize those descriptions.

This allows you to add a new layer of intelligence to your analysis of multimodal data. You can find more information in our multimodal documentation.

Wrapping up

These are just a few of the exciting new features in BigQuery AI. We are committed to continuously improving our platform to provide you with the best tools for building the next generation of AI-powered solutions. We can't wait to see what you'll create. For more information, check out the BigQuery AI documentation.