• 

  Seedcore Workflow Capability

• 

  Seedcore Architecture

Inspiration

As AI systems move beyond routine automation, people are making more creative, ambiguous, and open-ended requests—especially in physical environments like hotels, factories, and robotics-enabled spaces.

Most AI demos stop at text generation. I wanted to explore what happens after reasoning: How do we safely turn intent into real-world execution, handle failure, and preserve intent over time?

SeedCore was built to answer that question—by treating reasoning not as a UI feature, but as infrastructure.

What it does

SeedCore turns Gemini 3 into a reasoning supervisor for the physical world.

Instead of issuing opaque commands, Gemini 3:

• Interprets creative, multimodal intent
• Decomposes goals into verifiable task graphs (DAGs)

SeedCore then:

• Enforces policy-first, deny-by-default execution
• Routes tasks to authorized agents and tools
• Executes actions in real systems
• Feeds telemetry back for dynamic replanning

Intent → Plan → Policy → Execution → Feedback → Replanning

Gemini 3 reasons. SeedCore executes. Reality pushes back.

How we built it

SeedCore is a brand-new system, built specifically for this hackathon using the Gemini 3 API.

Key architectural components:

• Gemini 3 as a reasoning and planning supervisor (not a chatbot)
• TaskPayload v2.5+, a versioned execution envelope separating cognition, routing, and execution
• Policy Knowledge Graph (PKG) for pre-execution authorization
• PlanExecutor to materialize task graphs into executable child tasks
• Ray Actors & Ray Serve for distributed orchestration and isolation
• Holon Memory to preserve intent across failures and time

I built the system end-to-end as a solo developer. AI assistants were used as collaborators, but all architectural and implementation decisions were made by me.

Challenges we ran into

1. Preventing Gemini 3 from directly controlling tools or devices
2. Avoiding hallucinated or physically impossible actions
3. Handling creative, non-deterministic requests without slowing routine tasks
4. Preserving user intent when physical execution fails
5. Coordinating distributed agents safely with RBAC and provenance tracking
6. Designing schemas that evolve without constant migrations
7. Making deep infrastructure understandable in a 3-minute demo

Accomplishments that we're proud of

1. Built a closed-loop reasoning → execution → feedback system
2. Enabled high-surprise detection to selectively trigger deep reasoning
3. Introduced TaskPayload v2.5 as a stable execution contract
4. Converted natural language into executable DAGs
5. Enforced policy-first, deny-by-default execution
6. Implemented just-in-time agent and capability materialization
7. Unified multimodal reasoning with physical execution
8. Preserved intent through failure with Holon Memory
9. Scaled execution using distributed systems primitives
10. Demonstrated what “industrialized intelligence” looks like in practice

What we learned

• Reasoning models are most powerful when decoupled from execution
• Safety must be enforced before execution, not after
• Creative requests require different handling than routine automation
• Feedback loops are essential for real-world AI systems
• Infrastructure matters as much as intelligence

What’s next for SeedCore

1. Expand condition handling and long-running task orchestration
2. Add richer environment telemetry for smarter replanning
3. Support additional reasoning models beyond Gemini 3
4. Harden policy tooling for regulated environments
5. Explore more real-world deployments (hospitality, robotics, manufacturing)

Demo & Submission Notes (Gemini 3)

This is a brand-new application, created specifically for this hackathon.

• Gemini 3 is used as a reasoning engine, not a chat interface
• The demo video (under 3 minutes) shows:
  • Creative guest-style prompts
  • Gemini 3 task decomposition
  • Policy-checked execution
  • Feedback-driven behavior

Public links are provided to:

• Gemini API / AI Studio project
• Code repository (where applicable)

About the Builder

I’m a technology pioneer with experience as:

• Master’s degree from UESTC (University of Electronic Science and Technology of China)
• Senior technical leader at Motorola
• Senior architect at Alibaba
• Co-founder(CTO) of an IoT startup
• CTO of an AI manufacturing company

I’m a science-fiction enthusiast with strong self-discipline and an independent engineering mindset, focused on building systems that bridge imagination and reality.

Built With

• closed-loop
• distributed-agents
• gemini-3
• policy-enforcement
• ray
• ray-serve