GlassBox

GIF
Landing Page
Key Features
Decision Log

Inspiration

Customers often feel unsupported after making a purchase; onboarding to new platforms can be confusing, human support is slow or limited, and churn occurs because expectations aren’t met. While companies want to keep human intervention, they struggle to scale. As a result, most customer success teams simply can’t provide 24/7, personalized, proactive support, leading to missed opportunities and preventable churn.

Problems with Existing Solutions

Limited Multi-Modal Outputs: Most platforms are text-first, with limited support for video, images, PDFs, or other media required for onboarding & support.
Poor handling of uncertainty: Agents act even when confidence is low, leading to hallucinations or broken workflows
Lack of Transparency & Explainability: Most agentic platforms are black boxes. Teams can’t see why an agent fails, and the reasoning behind a certain decision is hidden. This makes it difficult to debug, audit, or trust the system.

Solution

GlassBox helps companies deliver scalable, personalized onboarding and support:

AI Generated Videos: Automated onboarding tutorials and use case demos tailored to each customer’s journey
AI Created Blogs & Articles: Guides, walkthroughs, and support content with images and text
AI Responses to Messages, Emails, and FAQs: Fast, accurate answers to customer questions. Escalation to human advisors when needed.

How we built it

Our platform helps non-technical teams build and deploy AI agents for their customer support strategies. We built a system that enables agents to work with unstructured documents (PDFs, CSVs, etc.) by converting them into a SQL database enriched with embeddings. This approach allows the agent to query, rank, and filter knowledge bases that don’t rely on third-party APIs.

For reasoning and interaction, we used Gemini as the core model. We extended it with modular “blocks” that let users define tools through simple API endpoints, which the agent can then call directly.

On the infrastructure side, we automated environment setup using the DigitalOcean API and GitHub Actions, which provisioned virtual machines and handled CI/CD pipelines to deploy and run the agents reliably.

Challenges we ran into

We initially struggled with designing the infrastructure to support AI agents. Unlike traditional applications, agents require a very specific setup: persistent storage for knowledge bases, APIs to expose tools, CI/CD for fast iteration, and compute that can scale with demand. Scaling was particularly challenging because agents aren’t just about throughput — they’re highly dependent on prompting quality and orchestration, which makes infrastructure harder to generalize.

A big problem was learning how to convert a lot of unstructured data quickly into SQL data tables that we can use for inferencing or embedding. Using databricks helped us easily automate that problem of convert unstructured data into structured data and running it for high intensity ML applications as well.

Another challenge we faced was scoping the problem. Our initial vision was to create a transparent way to build, deploy, and manage AI agents. While we had a clear use case for the demo, the platform itself could apply broadly across industries and use cases. That raised an important question: what truly sets us apart from other general-purpose agent platforms? Although we provide better traceability through step-by-step reasoning logs, clearer visibility into why each agent makes a decision, and stronger infrastructure and user experience, it wasn't enough. To sharpen our focus, we decided to target a specific audience. Through research and conversations, we discovered that many companies struggle with customer onboarding and retention, an area where our platform could deliver outsized impact. By helping companies build videos, blogs/articles and respond to inquiries, we can help customers who often feel unsupported after making a purchase.

Accomplishments that we're proud of

We are proud of building an API that helps others deploy their APIs LMAO! We’re also proud of connecting such a complex system end-to-end. Building an API that helps others deploy, manage, and orchestrate their APIs touches multiple layers of infrastructure, CI/CD, and automation, it’s a difficult problem, and successfully integrating it into a coherent workflow was a major achievement.

What we learned

Through this project, we learned a great deal about system design and distributed systems, particularly by acting as system administrators for deploying APIs on behalf of others through our backend. We gained hands-on experience with AI agents and the best practices for building infrastructure to support them, including creating a memory layer, implementing logging for data compliance and governance, and integrating a layer for querying our Databricks knowledge base.

This project was also one of our first experiences using Databricks, working with large datasets, and performing inference at scale, which made it an incredibly fun and educational experience. Overall, it gave us a deep understanding of both the technical and operational challenges involved in building scalable, intelligent systems.

What's next for GlassBox

AI Workflow Builder: Users can input a request → AI auto-generates the right workflow
Expanded Multi-Channel Output: Expand AI outputs across multiple channels, including call support and social media
Full Customer Journey Coverage: Cover the whole journey from onboarding → adoption → retention. Target non-users with AI-generated social posts and outreach campaigns.