Why I Left Web3 to Build AI: A Founder's Journey

The Moment I Knew

It was November 2023. I was deep in the Web3 world -- building full-stack applications at MagicSquare, a Web3 app store that had onboarded over 190,000 users. The product was live, growing, and technically sophisticated. We had built token-gated experiences, decentralized identity flows, and on-chain analytics dashboards. I should have been thrilled.

Instead, I found myself spending every evening reading papers about transformer architectures and watching Andrej Karpathy's neural network lectures. Not because my job required it, but because I could not stop thinking about what this technology would mean for the world. The Web3 products I was building solved real problems for a niche audience. The AI systems I was reading about had the potential to change how billions of people accessed information, education, and services.

That tension between what I was building and what I wanted to build became unbearable. So I made the leap.

The Web3 Years: What I Built and What I Learned

My Web3 journey was not a detour -- it was a masterclass in building products under extreme constraints. At MagicSquare, I was a full-stack engineer responsible for both the consumer-facing app store and the backend infrastructure. At LunarCrush, I worked on social analytics for crypto assets, processing millions of social media signals in real time to generate sentiment scores.

The technical challenges were genuinely interesting. Blockchain applications demand a level of correctness that most web applications can skip. When your code handles financial transactions on an immutable ledger, there is no "we'll fix it in the next deploy." I learned to write exhaustive tests, to think adversarially about edge cases, and to design systems where failure modes were explicit and recoverable.

# A pattern from my Web3 days that I still use in AI systems:
# Explicit state machines with auditable transitions
class TransactionState:
    PENDING = "pending"
    VALIDATING = "validating"
    EXECUTING = "executing"
    CONFIRMED = "confirmed"
    FAILED = "failed"
    ROLLED_BACK = "rolled_back"

    VALID_TRANSITIONS = {
        PENDING: [VALIDATING, FAILED],
        VALIDATING: [EXECUTING, FAILED],
        EXECUTING: [CONFIRMED, FAILED],
        FAILED: [ROLLED_BACK],
    }

    @classmethod
    def validate_transition(cls, current: str, target: str) -> bool:
        allowed = cls.VALID_TRANSITIONS.get(current, [])
        if target not in allowed:
            raise InvalidTransitionError(
                f"Cannot transition from {current} to {target}. "
                f"Allowed: {allowed}"
            )
        return True

This pattern -- explicit state machines with validated transitions -- has been invaluable in building AI agent systems where you need to track and audit every step of a complex workflow.

The Decision to Leave

Three observations pushed me toward AI:

The impact ceiling. Web3 products, at least in 2023, served a relatively small audience. The total addressable market of crypto-native users willing to engage with decentralized applications was in the low millions globally. AI, meanwhile, was poised to transform how every person interacted with technology. I wanted to work on problems with a larger blast radius.

The builder's dilemma. In Web3, a significant portion of engineering effort goes toward the infrastructure layer -- wallets, gas optimization, cross-chain bridges -- rather than user-facing value. I spent more time working around blockchain limitations than building features users loved. In AI, the infrastructure was maturing rapidly (thanks to OpenAI, Anthropic, and the open-source community), which meant I could focus on the application layer.

The personal calling. I grew up in Nepal. I watched my family navigate government bureaucracies that were inaccessible, confusing, and exclusively in languages that excluded large portions of the population. When I saw what voice AI and multilingual NLP could do, I did not see a technical curiosity -- I saw a solution to a problem I had lived with my entire life.

The Transition: Harder Than Expected

Leaving Web3 was not as simple as updating my LinkedIn title. The AI engineering landscape in late 2023 was simultaneously exciting and chaotic. Everyone was building "AI-powered" products, but few had production-grade engineering practices. The gap between a ChatGPT wrapper and a reliable AI system was enormous, and I needed to bridge it quickly.

I joined Fundora, a personalized investment advisory platform, as an AI engineer. This was deliberate -- it let me apply my fintech knowledge from Web3 while learning the AI stack in a production context. At Fundora, I built my first reinforcement learning system, my first RAG pipeline, and my first model evaluation framework. The learning curve was steep, but my Web3 background in building robust, auditable systems gave me an unexpected advantage.

Most AI engineers at the time came from research backgrounds and treated production concerns as afterthoughts. I came from a production engineering background and treated model performance as the concern. The combination turned out to be exactly what AI products needed: someone who cared equally about model accuracy and system reliability.

Founding Sarathi Studio

The work at Fundora taught me the craft of AI engineering. But the spark that became Sarathi Studio came from a conversation with a government official in Assam during a visit home. He described the challenge of delivering digital services to citizens who spoke Assamese, Bodo, or other indigenous languages. "We have the services," he said. "We don't have the interface."

That conversation crystallized everything. I had the technical skills to build multilingual voice AI. I had the product sensibility from years of building user-facing applications. And I had a deeply personal motivation. Sarathi Studio was founded to build AI systems that serve populations traditionally excluded from the digital revolution.

The name Sarathi comes from Sanskrit -- it means "charioteer" or "guide." In the Mahabharata, Lord Krishna served as Arjuna's sarathi. Our AI serves as a guide for citizens navigating complex systems.

Joining AlysAI: The Enterprise Dimension

While building Sarathi Studio, I joined AlysAI as an AI Engineer. AlysAI works on enterprise AI -- vision-aware agents, real estate automation, compliance systems. This dual role (startup founder and enterprise engineer) has been the most productive arrangement of my career.

At AlysAI, I learn how enterprise clients think about AI: reliability, auditability, cost predictability, and integration with existing systems. These lessons flow directly into how I build Sarathi's products. At Sarathi Studio, I push the boundaries of what is possible with low-resource languages and constrained deployment environments. Those innovations flow back into my AlysAI work as novel approaches to difficult problems.

Lessons for Others Considering the Leap

Your previous domain is an asset, not baggage. Every industry I have worked in -- Web3, fintech, government tech -- has informed how I build AI systems. Domain expertise combined with AI skills is far more valuable than AI skills alone.

Start building before you feel ready. I did not wait until I had "mastered" AI to start building. My first RAG system was embarrassingly naive. My first model evaluation pipeline had bugs. But each project taught me things no course or paper could.

Choose problems you care about personally. The transition from Web3 to AI was hard. What made it sustainable was not the career opportunity -- it was the conviction that voice AI could genuinely improve people's lives. That conviction carries you through the late nights, the failed experiments, and the moments when you question whether you made the right choice.

# My career transition as code (tongue in cheek)
career = Pipeline([
    Stage("web3", skills=["full_stack", "blockchain", "fintech"]),
    Stage("transition", skills=["ml_fundamentals", "nlp", "rl"]),
    Stage("ai_engineering", skills=["rag", "voice_ai", "mlops"]),
    Stage("founder", skills=["product", "team", "vision"]),
])

# The key insight: it is not a linear path, it is a DAG
# Each stage feeds into all subsequent stages
for stage in career.stages:
    stage.connect_to(career.future_stages(stage))

Where I Am Now

Today, I split my time between AlysAI (building enterprise AI agents) and Sarathi Studio (building voice AI for underserved populations). The contrast keeps me sharp. Enterprise work teaches me rigor. Social impact work teaches me empathy. Together, they make me a better engineer than either alone.

If you are considering a similar transition -- from Web3, from traditional software engineering, from any other field into AI -- my advice is simple. The world does not need more AI researchers publishing papers. It needs more engineers who can build AI systems that work reliably for real people, solving real problems. If that describes you, the AI field needs you. Make the leap.