Building Sofi: Exploring a More Local and Accessible Future for AI 🤖

Ilakkiyan J — Fri, 21 Aug 2026 09:03:48 +0000

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.

Building AI Systems, One Project at a Time 🚀

Ilakkiyan J — Fri, 21 Aug 2026 08:53:13 +0000

Hey DEV Community! 👋

I’m Ilakkiyan, a Software Engineer and AI Builder interested in the space where software engineering meets artificial intelligence.

I’ve always enjoyed building things — not just following tutorials, but taking an idea, figuring out how it should work, and turning it into a real system.

That journey has led me from traditional full-stack development into AI engineering, LLMs, AI agents, and intelligent applications.

🧠 What I’m Exploring

Right now, I’m particularly interested in:

🤖 AI Agents and Agentic AI
🧠 Large Language Models (LLMs)
🔎 RAG and knowledge-based AI systems
⚙️ Backend engineering and APIs
💻 Full-stack application development
🖥️ Local and offline AI
🚀 Building AI products that solve real problems

One thing I’ve learned is that building AI applications isn't simply about calling an LLM API.

The interesting engineering problems are often around data, architecture, reliability, tool usage, context, evaluation, security, and user experience.

That's where I want to spend more time.

🛠️ What I've Been Building

One of my recent projects is Sofi, an offline AI desktop assistant.

The goal was to explore what happens when an AI assistant doesn't depend entirely on cloud APIs.

The system combines technologies such as:

React
Electron
FastAPI
Python
Ollama
ChromaDB
Vosk
Coqui TTS

Working on projects like this has made me more interested in the engineering behind AI systems — especially how different components come together to create something that actually feels like a product.

I've also worked on Medorc, a secure healthcare platform focused on real-time health data orchestration and decision support.

Each project has taught me something different.

📚 Why I'm Joining DEV.to

I'm joining DEV.to because I want to learn in public.

Instead of keeping everything inside my laptop, I want to document:

What I'm building
What breaks
What I learn
Engineering decisions I make
AI experiments
Interesting problems I encounter
Things I wish I had understood earlier

Some posts will probably be polished.

Some might be experiments.

And some will probably start with:

“I had no idea what I was doing, so I tried this…”

That's part of the process.

🚀 What's Next?

I'm planning to spend the coming months going deeper into AI engineering and agentic systems, while continuing to strengthen my foundations in software engineering.

I'll be sharing the journey here — including projects, experiments, technical deep dives, failures, and lessons learned.

If you're also building with AI, LLMs, agents, or full-stack technologies, I'd love to connect and learn from your work.

This is Day 1.