Hackathons are often associated with sleepless nights, countless cups of coffee, and a frantic sprint to the finish line. But my experience at the Gemini Live Hackathon was about much more than just shipping code—it was about building something genuinely innovative, mentoring the next generation of developers, and extending that impact into the open-source world.

StoryForge: A Live AI Storytelling Agent

Our team of five set out to build StoryForge—an interactive, real-time storytelling agent that you could speak to. The core idea was simple but ambitious: a user starts a conversation, and the AI dynamically generates a rich, illustrated story in real time. At any point, the user can steer the narrative in a completely new direction just by talking.

The Architecture

Under the hood, StoryForge was powered entirely by the Google ecosystem:

• Google ADK (Agent Development Kit) served as the central orchestration layer for our live agent.
• Gemini Live Models handled the real-time voice conversation with the user.
• Gemini 2.5 Flash Preview was used as a tool by the live agent, responsible for generating the story text dynamically on each turn.
• Imagen 4 was called as a second tool to generate vivid illustrations for each scene on the fly.

One of the key architectural challenges we tackled was context management. A live conversational agent has a limited context window, and you absolutely cannot fill it with an ever-growing story. Our solution was a metadata-snippet architecture: instead of feeding the full narrative back into the live agent, we generated concise metadata summaries of the story’s current state—characters, plot points, emotional tone—and fed those to the agent. This kept the live agent lightweight and responsive while still maintaining perfect narrative coherence, even across long, winding stories.

Mentoring Aman: Agentic AI Observability

Beyond the project itself, the hackathon became a platform for mentorship. I had the privilege of working closely with Aman, a 2nd-year IT bachelor’s student, throughout the event. Together, we focused on one of the most critical—and often overlooked—aspects of agentic AI systems: observability.

When AI agents make autonomous decisions, chain tool calls, and dynamically route between models, you need to see what’s happening inside the pipeline. We built an observability pipeline for StoryForge using Google’s tooling, covering:

• Agent state tracking — monitoring how the live agent’s internal state evolved across conversation turns.
• Tool-calling telemetry — logging every invocation of Gemini Flash and Imagen 4, including latency, token usage, and response metadata.
• Pipeline tracing — end-to-end visibility into the flow from user speech → live agent → story generation → image generation → response.

Watching Aman go from grappling with these concepts to confidently instrumenting the pipeline was one of the most rewarding parts of the entire hackathon.

From Hackathon to Open Source: Aman’s First Matplotlib Contribution

The mentorship didn’t end when the hackathon wrapped up. I firmly believe that contributing to open source is one of the best accelerators for a developer’s growth, and with Google Summer of Code (GSoC) 2026 on the horizon, I wanted to help Aman build a track record of meaningful contributions.

We targeted Matplotlib—one of the most widely-used scientific visualization libraries in the Python ecosystem, with over 20k GitHub stars and a massive, active community. Navigating a codebase of that scale can be intimidating for any developer, let alone a 2nd-year student.

We worked through the process together:

1. Finding the right issue — We identified Issue #17479, a long-standing documentation gap where the multipage_pdf gallery example had no thumbnail because it uses pdf.savefig() instead of plt.savefig(), preventing Sphinx Gallery from auto-generating one.
2. Understanding the fix — The solution required creating a custom static SVG thumbnail and wiring it up via the sphinx_gallery_thumbnail_path directive.
3. Crafting the PR — Aman wrote a clean, well-documented PR with a clear problem statement, root cause analysis, and fix description.

The result? PR #31308 was merged into Matplotlib’s main branch. 🎉

Seeing that purple “Merged” badge appear was a massive milestone for Aman, and it sets a strong foundation for his GSoC 2026 application. Since then, we’ve continued collaborating on additional open-source contributions together, and the growth I’ve seen in just a few weeks has been incredible.

Why Mentorship Matters

The Gemini Live Hackathon gave our team of five a platform to build a cutting-edge agentic AI application, but the impact went far beyond the event itself. It created a mentorship pipeline that connected real-world AI engineering with open-source contribution and career development.

If you’re an experienced developer, I’d encourage you to take a junior developer or student under your wing. Find an open-source project, pair through a tough issue, and help them land that first PR. The compound returns—for them, for you, and for the community—are immeasurable.