SilverAgent | Devpost

Home Page
Booking Ride 1
Booking Ride 2
SEA-LION Thinking
SeaLION-TC-v1 fine tuning

Inspiration

Our team, "Team Refused Bequest", identified a critical problem in Singapore's digital landscape: "App Fatigue". With a multitude of apps like HealthHub, LifeSG, and DBS PayLah, the ecosystem has become fragmented. We found that while a typical phone might have 80 apps, a user only frequently uses about 9 of them. This complexity affects over 1 million Singaporeans, particularly seniors who struggle to navigate these diverse interfaces. We wanted to create "The Digital Grandchild Every Senior Deserves", a proactive agent that serves as a unified interface so seniors don't have to juggle multiple apps.

What it does

SilverAgent is a sovereign agentic AI designed to be the proactive agent of the future. It simplifies digital interactions by providing a single interface for various services.

Multilingual Support: It accepts voice notes in up to 13 South East Asian languages.
Contextual Understanding: It understands intent, such as knowing that "bone doctor" refers to an Orthopaedic specialist.
Action-Oriented: Unlike standard chatbots that just provide information, SilverAgent takes action. It can book appointments or rides across systems in under 60 seconds.
Family Inclusion: It allows for family involvement in healthcare decisions.
Zero Learning Curve: It is designed to meet seniors on an interface they are already used to.

How we built it

We utilized an "Edge-Native Architecture" to prioritize privacy and performance.

The Brain (SeaLION-TC-v1): We fine-tuned the Qwen SEA-LION v4 8B VL model using QLoRA to enable complex agentic tool calling. This allows the model to understand user intent in local languages.
The Protocol (MCP): We used the Model Context Protocol (MCP) as the infrastructure to orchestrate modular components. This acts as the "connective tissue" between different app services.
Deployment: The model is deployed as a Quantised Q4_K_M model, which enables high-performance, private, and offline-capable inference on standard hardware.

Challenges we ran into

Broad Context Issues: Generalist models like ChatGPT often lack specific context for Singapore or fail to access specific features because their scope is too broad.
Fragility of Existing Bots: DIY bots are often brittle and limited to single systems, while standard chatbots provide FAQs but cannot execute bookings.
Privacy Concerns: To protect vulnerable seniors, we needed to ensure data stayed local rather than relying solely on cloud processing.

Accomplishments that we're proud of

Performance Uplift: Our fine-tuned model, SeaLION-TC-v1, achieved a score of 75.00 in Live Parallel (Multi-Tasking) metrics, up from 50.00 in the base model.
SOTA Compliance: We achieved a +25% uplift in Parallel Execution and +12% in Relevance on the BFCL v4 benchmark.
Competitive Capability: Our model rivals OpenAI's 4o model in complex tool calling tasks.
First Mover Advantage: We are the first in Singapore to automate these common services, creating a defensible moat.

What we learned

Human-Centric Programming: We realised that the new programming language is "human". Instead of navigating apps, users should just be able to express their intent.
Unlocking Smart Nation: We learned that we didn't need to rebuild the Smart Nation infrastructure; we just needed to unlock it using MCP as a unified interface.

What's next for SilverAgent

We have laid out a comprehensive 1-year roadmap to scale to a national level.

Pilot Program: A 3-month trial targeting 1 million+ users to collect data and feedback on frequently used apps.
Partnerships: We plan to collaborate with GovTech Singapore, IMDA, and Lion Befrienders.
Technological Expansion: We aim to deploy an API for third-party integration and explore "Mixture-of-Agents" (MOA) where swarms of autonomous agents handle daily menial tasks simultaneously.