TL;DR
Got selected out of 4,000+ applicants for the Google DeepMind Bangalore Hackathon 2026 (250 builders made it in). My teammate and I merged two hackathon tracks — Conversational Video & Motion with Omni Flash and Local-First Agents with Gemma 4 — into one pipeline: a multi-turn conversational video orchestration agent that asks clarifying questions and iteratively refines its output, backed by on-device reasoning for state management. No trophy this time, but a working prototype and a lot of lessons. Here's the build.
The selection
Google DeepMind and Cerebral Valley ran this hackathon in Bangalore, and the numbers alone made it feel different from the college-circuit events I'd been doing: 4,000+ applications, 250 seats. Getting the acceptance email didn't fully register until I was actually in the room — this wasn't a campus event anymore, it was a room full of people who build for a living.
The idea: merging two tracks instead of picking one
The hackathon offered several tracks. Two stood out to me and Rohan:
- Conversational Video & Motion with Omni Flash — generate and iterate on video through natural conversation
- Local-First Agents with Gemma 4 — on-device reasoning and autonomous decision-making
Instead of choosing one, we asked: what if the video generation system was the agent? So we built a pipeline where:
- Gemini Omni Flash + NB2 Lite handle multi-turn conversational video orchestration — the system doesn't just take a prompt and spit out a video. It asks clarifying questions, holds context across turns, and iteratively refines the output based on what the user actually meant.
- Gemma 4 runs on-device for state management and decision-making — tracking where the conversation is, what's already been generated, what still needs clarifying, and deciding the next action locally instead of round-tripping everything through a heavier model.
The result was an agent that blends reasoning, local intelligence, and multimodal generation into a single loop, rather than three separate systems bolted together.
What building under a deadline like this actually felt like
Tight deadlines are a hackathon constant, but the density of the room changes the pressure. Every table around us was shipping something legitimately interesting, which is equal parts motivating and intimidating. Having a partner who's good at both brainstorming and debugging (thanks, Rohan) mattered more here than in any hackathon I'd done before — there wasn't time to context-switch between "figuring out the architecture" and "fixing why the state manager is out of sync."
The outcome
We didn't place. No trophy this round. But we left with:
- A working end-to-end prototype of the conversational video + local agent pipeline
- Direct exposure to how a team like DeepMind's frames the "agentic + multimodal" problem space
- Genuinely useful conversations with other builders working on adjacent problems
- Insights from Amit Vadi and Ray Del Vecchio, who spoke at the event
Why I'm writing this up
Hackathons are noisy — you build for 24-48 hours, present, and the artifact usually dies in a GitHub repo somewhere. Writing it up is partly documentation for myself (what worked, what I'd do differently) and partly a way of tracking how the kind of problems I'm choosing to work on has shifted — from single-model chatbots to multi-agent, multimodal, local-first systems.
If you're building anything similar — conversational video generation, on-device agent state management, or just curious about how Omni Flash and Gemma 4 play together — I'd genuinely like to compare notes. Drop a comment or find me on GitHub.
This is part of an ongoing hackathon log — you can find the earlier entries (including two first-place wins and a near-miss with a legal-aid kiosk built on a multi-agent swarm) on my blog.
Top comments (0)