FlashDraft

A local-first AI creative director. Gemma 4 runs entirely on-device as the "brain," deciding when to hand off to cloud-based generation — NB2 Lite for instant images, Omni Flash for conversational video. Built for the Google DeepMind Bangalore Hackathon.

---

The problem

AI video production tools today force a bad trade-off:

• Every iteration costs money. Fix a typo, tweak a prompt — you're burning cloud API credits even during brainstorming.
• They break the moment your connection does. No offline recovery, no resume — you lose your session.
• Your creative IP leaves your machine. Scripts, plot twists, unreleased campaigns — all sent to a third-party server just to draft an idea.

FlashDraft fixes this by keeping the thinking local and only paying for the cloud when it counts.

Who it's for

• Filmmakers & directors blocking out scenes before committing to expensive production or VFX.
• Location scouts working in the field with unreliable connectivity, visualizing a site under different lighting/weather.
• Indie game devs & storyboard artists iterating on concept sequences without per-generation cloud costs eating their budget.

USP

A fault-tolerant, privacy-first AI director, not another API wrapper:

1. Cost-free iteration, premium rendering — brainstorming and storyboarding run locally and for free on Gemma 4. Expensive cloud calls (NB2 Lite, Omni Flash) only fire when the shot is ready.
2. Network-resilient continuity — if the connection drops mid-session, the local agent doesn't crash. It falls back to a local text storyboard and resumes rendering the moment connectivity returns.
3. IP isolation — the creative core (script, plot, character logic) never leaves the device. Only minimal, isolated visual instructions go to the cloud.

Our USP is a hybrid-agent architecture that puts a local Gemma 4 "Director" on your device for cost-free, private brainstorming and network resilience, while leveraging NB2 Lite and Omni Flash in the cloud only to render the final video.

Architecture

        ┌─────────────────────────────┐
        │   Gemma 4 E4B (on-device)   │
        │   sense → decide → act → check
        └──────────────┬──────────────┘
                        │
          ┌─────────────┴─────────────┐
          │                           │
   [Tool] NB2 Lite            [Tool] Omni Flash
   fast base image             conversational video
   generation (~4s)            edit + motion
          │                           │
          └─────────────┬─────────────┘
                         │
              network error caught?
                         │
                  ┌──────┴──────┐
                  │  local text  │
                  │  storyboard  │
                  │  fallback    │
                  └──────────────┘

Loop:

• Sense — Gemma 4 parses the user's request (text/voice) entirely offline.
• Decide — checks local state/history, decides which tool to call and what minimal instruction it needs.
• Act — dispatches to NB2 Lite (image) and/or Omni Flash (video/motion edit).
• Check — verifies the result; if a cloud call fails (or the network drops), catches it and falls back to a local, detailed storyboard instead of crashing, resuming the cloud render once connectivity returns.

Tech stack

Layer	Tool
Local orchestrator	Gemma 4 E4B, run via LM Studio (local OpenAI-compatible server) or Ollama
Agent framework	Antigravity SDK (Python)
Fast image generation	Nano Banana 2 Lite (NB2 Lite)
Conversational video	Gemini Omni Flash (gemini-omni-flash-preview)
Client	Python (openai SDK pointed at localhost)

4. Add cloud keys (for NB2 Lite / Omni Flash)

• Use a billing-disabled project for NB2 Lite (free tier, ~1,000–1,500 req/day).
• Use a separate, billing-enabled project for Omni Flash ($0.10/sec of video). Keeping these separate avoids losing your free tier on incidental calls once billing is enabled on a project.

Built for

Google DeepMind Bangalore Hackathon