Inspiration
We are entering an era of "Legacy Inertia." Despite the rapid advancement of AI, trillions of dollars in global commerce are still "glued" together by human operators manually moving data from unstructured sources—like physical documents and voice calls—into rigid software systems that lack modern APIs. We built Synapse Agentic Platform to be the autonomous bridge for this "last mile," transforming the human role from manual data-entry to high-level strategic oversight.
What it does
Synapse is a vertical-agnostic orchestration platform that utilizes Agentic Reasoning to automate complex, high-stakes workflows. It operates through a hierarchical multi-agent system:
The Coordinator: A high-level reasoning engine that deconstructs abstract intents into a verifiable Asynchronous Finite State Machine.
The UI Navigator: A specialized multimodal worker that "sees" and interacts with legacy interfaces in real-time, executing precise on-screen actions without requiring backend integrations.
Live Supervision: A bidirectional streaming interface (Gemini Live) that allows supervisors to monitor progress and "barge-in" via voice to guide the system during critical edge cases.
How we built it
Developed within Google Antigravity, Synapse leverages the ultra-low latency of Gemini Flash. The engine is built with the Google Agent Development Kit (ADK) and hosted on Google Cloud Run. We used Firestore as a grounded memory layer to maintain process state across long-running workflows. Our infrastructure is fully automated via Cloud Build, ensuring a robust CI/CD pipeline from the first line of code to production.
Challenges we ran into
Our biggest hurdle was ensuring the UI Navigator could maintain precision across dynamic legacy layouts while simultaneously processing real-time voice feedback. Synchronizing multimodal vision with bidi-streaming required deep optimization of our agentic planning cycles to prevent "hallucinated clicks" and ensure the agent remained grounded in the actual pixels on the screen.
Accomplishments that we're proud of
We successfully built a platform that can "automate the un-automatable." Witnessing the agent autonomously navigate a complex enterprise dashboard to fulfill a process derived from a physical document—without a single API call—proved that the "Digital Gap" is finally bridgeable.
What we learned
We learned that the future of AI isn't in chatbots, but in Agentic Workflows. The ability for an agent to have a "Live" conversation with a human while performing physical digital labor is a paradigm shift. We also discovered that Gemini Flash's multimodal speed is the secret sauce for making these real-time interactions feel natural rather than robotic.
What's next for Synapse Agentic Platform
We aim to expand Synapse into a cross-platform "Universal Hand," moving beyond the browser to navigate native desktop and mobile applications. We are also exploring Long-Term Multimodal Memory (via Vertex AI Search) to allow Synapse to learn a company's unique "Business Intuition" over time, making it more autonomous with every process it completes.
Log in or sign up for Devpost to join the conversation.