Meet the Builders — Asia Pacific
AI is evolving incredibly fast.
But as I started building with AI, I kept coming back to a simple question:
What if an AI assistant could do more while keeping its intelligence and data closer to the user?
That question became the starting point for Sofi, an AI desktop assistant I built to explore the possibilities of local AI, desktop automation, semantic memory, and voice interaction.
For me, Sofi isn't just another chatbot.
It's an experiment in understanding what happens when we treat an AI assistant as a complete software system rather than simply connecting an application to an LLM API.
🌏 Why Local AI?
A lot of modern AI applications depend heavily on cloud infrastructure.
That provides incredible capabilities, but it can also introduce challenges around:
- Internet dependency
- Data privacy
- API availability
- Recurring API costs
- Latency
- Control over personal data
As a developer from India, I’m particularly interested in building technology that can work across different infrastructure and connectivity environments.
That led me to explore a different approach:
Can useful AI capabilities run directly on a user's machine?
🧠 Introducing Sofi
Sofi is a fully offline AI desktop assistant built around local LLM inference.
The system combines:
- React for the interface
- Electron for the desktop application
- FastAPI for backend services
- Ollama for local LLM inference
- ChromaDB for semantic memory
- Vosk for voice recognition
- Coqui TTS for voice responses
But the interesting part isn't just the technology stack.
Sofi integrates 20+ desktop automation tools that allow the assistant to interact with the user's computer.
Instead of only answering:
“What should I do?”
the assistant can move toward:
“Let me do it for you.”
⚙️ From Chatbot to Agent
One of the biggest parts of Sofi was developing an AI tool-calling framework.
The assistant can use tools for tasks such as:
- Desktop automation
- File management
- Application control
- Web search
- Windows system operations
This changed the way I thought about AI applications.
A traditional chatbot follows a relatively simple flow:
User → Prompt → Model → Response
An agentic system can instead follow:
User → Understand Goal → Reason → Select Tool → Execute → Observe → Continue
That creates an entirely different engineering problem.
Now the application needs to understand:
- Which tool should be used?
- What arguments does it require?
- Is the action safe?
- What happened after execution?
- Should another tool be called?
- How should the result be presented to the user?
The LLM becomes one component inside a much larger system.
🔐 Privacy as an Engineering Consideration
Another reason I wanted to explore local AI was privacy.
An AI assistant can potentially interact with:
- Personal files
- Documents
- Conversations
- Voice input
- Application data
- User preferences
Keeping processing local where possible can reduce the amount of information that needs to leave the user's machine.
That doesn't automatically make an application completely private or secure.
There are still important challenges around:
- Local data protection
- Permissions
- Authentication
- Tool execution safety
- Prompt injection
- Sensitive information handling
But these considerations become part of the architecture from the beginning.
🧩 What Building Sofi Taught Me
1. AI engineering is systems engineering
Before working deeply with AI, it was easy to think of AI development as primarily:
Prompt + Model = Application
Building Sofi changed that perspective.
The model is only one component.
The real system includes:
UI → Backend → Model → Memory → Tools → Execution → Feedback
Every layer introduces engineering challenges.
2. Agentic AI requires controlled execution
Giving an AI access to tools is powerful.
It is also dangerous if done without proper boundaries.
An agent that can modify files or interact with the operating system needs careful control over what it can execute.
This made me think more about permissions, validation, tool boundaries, and predictable execution.
3. Local models introduce interesting trade-offs
Running models locally provides greater control and can improve privacy and offline availability.
But it also means working within the limits of the user's hardware.
That creates an interesting engineering trade-off between:
Model capability ↔ Latency ↔ Hardware ↔ Privacy ↔ Cost
There isn't one perfect architecture.
The right architecture depends on the problem.
🚀 My Broader AI Journey
Sofi is part of a larger journey I'm taking into AI engineering and Agentic AI.
I'm currently working as an AI Intern through the AICTE–IBM SkillsBuild–1M1B Applied AI program, where I've been working with prompt engineering patterns, IBM Granite models, Agentic AI, and RAG pipelines for sustainability use cases.
I've also built Medorc, an AI-powered healthcare platform.
Medorc includes 50+ type-safe REST APIs, JWT-based role access control, and a RASA-based healthcare assistant with 20+ intents and 10+ custom entities.
These projects have given me exposure to different sides of AI engineering:
Local AI → Agentic AI → RAG → NLP → Backend Systems → Real-world Applications
And I'm still learning.
🇮🇳 Building From India
One of the things that excites me most about the current AI ecosystem is that powerful tools are becoming accessible to builders everywhere.
You don't need to be part of a huge research lab to experiment.
A student or independent developer can take an idea, build a prototype, test it, share it, and learn from people across the world.
That's especially exciting in Asia Pacific, where different countries and communities have very different needs, languages, infrastructure, and constraints.
The opportunity isn't simply to build another AI demo.
It's to build AI that fits the realities of the people using it.
🌱 What's Next for Sofi?
Sofi is still an evolving project.
Some areas I want to explore next include:
- More capable autonomous agents
- Better long-term semantic memory
- Multimodal interaction
- Improved voice interfaces
- Safer tool execution
- Better local model routing
- More intelligent desktop workflows
- Hybrid local + cloud AI architectures
The goal isn't simply:
“Make Sofi smarter.”
It's:
“Make Sofi genuinely useful.”
🙌 Why I'm Sharing This
I'm participating in Meet the Builders because I believe builders should document the problems they're trying to solve—not just the final product.
There are thousands of developers across Asia Pacific experimenting with AI in ways that deserve more attention.
Some are building for agriculture.
Some are working on healthcare.
Some are solving language barriers.
Some are improving education.
And some are simply experimenting with a question that might eventually become something much bigger.
Sofi started with one of those questions:
What could an AI assistant look like if it could think, remember, speak, and act locally?
I don't have all the answers yet.
But I'm building my way toward them.
Build → Break → Learn → Build Better. 🚀
🔗 Let's Connect
I'm documenting my journey through AI engineering, Agentic AI, full-stack development, and practical AI systems.
If you're building something interesting somewhere in Asia Pacific, I'd love to hear about it.
We're all building the future from different starting points.
Let's share what we're learning along the way.
Top comments (0)