In late 2023, we made a decision that would consume most of our team's bandwidth for the next ten months: we were going to build an AI agent platform unlike anything in the market, one that lived simultaneously on Roblox, across the open web, and inside Web3 ecosystems. An agent that wasn't just a chatbot, but an AI KOL, market analyst, and fully functional NPC inside decentralised games. Powered by DeepSeek R1. Built on Base and Solana.Blogging is a powerful tool for businesses looking to establish their online presence and connect with their audience. A well-maintained blog can drive traffic, improve SEO, and enhance brand awareness.

The result is GiantAI, the first AI Agent platform simultaneously live on Roblox, Web2 and Web3. Both the platform app (friendly-giant-ai-app) and the smart contracts (friendly-giant-contracts) are open source on GitHub. It works. We're proud of it. And we made every mistake you'd expect a team to make when building something genuinely new in a space evolving faster than we could read the documentation.

This is that story, told as honestly as we can tell it because the lessons are more useful to the people building AI products today than another polished case study.

What GiantAI Actually Is

GiantAI is not a chatbot. It's a multi-role AI agent platform with a scope that, when we first sketched it on a whiteboard, made at least one team member visibly uncomfortable.

The platform operates across three distinct environments simultaneously:

• Roblox, as an AI NPC agent that lives inside decentralised game worlds, interacting with players, analysing in-game economies, and acting as a living character with persistent memory and personality
• Web2, as an AI KOL (Key Opinion Leader) and market analyst, surfacing insights, commentary, and on-chain data to audiences across social and web platforms
• Web3 / Base & Solana, deeply integrated with both Base and Solana blockchains, with open-source smart contracts and on-chain capabilities that make GiantAI a native participant across multiple decentralised ecosystems

The reasoning backbone is DeepSeek R1, chosen for its exceptional performance on analytical and reasoning tasks, its cost efficiency at scale, and its suitability for the kind of structured, multi-step thinking that an agent acting across three different environments requires.

The core ambition: an AI agent that isn't just deployed in Web3, it understands Base and Solana natively, participates in on-chain activity, and can act autonomously across both ecosystems.

That sounds compelling in a pitch deck. Building it took 10 months and a fundamental rethink of how AI agents should be architected when the environment itself is non-deterministic.

The Architecture We Started With (And Why We Threw It Away)

Our first architecture was what you'd call naive-monolith. A single backend service handled everything: model calls, tool execution, memory management, session state, and the embedding API. It worked in development. It fell apart at the first real load test.

The core problem: LLM calls, tool execution, and memory operations have completely different latency and reliability profiles. Bundling them into a single request-response cycle meant that a slow web search (which could take 8–12 seconds) blocked everything downstream — including the streaming response the user was waiting for.

The rewrite, which we should have done from the start, separated these into independent async services. The orchestrator dispatches work and assembles results. The LLM service handles model calls with streaming. The tool runner executes web searches, API calls, and code in isolated sandboxes. The memory service manages vector storage and key-value state. Each can scale, fail, and retry independently.

The Build Timeline

Why We Chose DeepSeek R1, Not GPT-4

This is the question we get most often at demos. The AI development world defaults to GPT-4 or Claude for agent reasoning. We went with DeepSeek R1. Here's why.

GiantAI operates in contexts that demand structured, step-by-step reasoning over complex data, on-chain analytics, game economy modelling, multi-step web research. DeepSeek R1 was built specifically for this kind of chain-of-thought reasoning, and in our internal benchmarks it consistently outperformed GPT-4 on tasks that required logical sequencing and analytical precision.

The other factor was economics. An AI agent that lives inside Roblox games and interacts with potentially thousands of players simultaneously has a very different cost profile than an enterprise chatbot handling 50 queries a day. DeepSeek R1's token costs made truly scalable deployment viable in a way that GPT-4 pricing simply wouldn't have allowed at our target scale.

We also valued the alignment with our multi-chain positioning. DeepSeek R1's open architecture made it possible to reason transparently about on-chain data, across both Base and Solana, in ways that matter for a platform where trust and verifiability are core to the user relationship.

The Mistakes (The Useful Part)

What We'd Do Differently

What GiantAI Taught Us About Building AI Products

The most important thing we learned is that building an AI product requires a team that has actually shipped AI in production before. The failure modes are different. The debugging process is different. The relationship between specification and behaviour is different.

The things that matter in traditional software like clean architecture, test coverage, type safety still matter. But they're table stakes. What separates good AI products from broken ones is prompt engineering craft, evaluation discipline, and a deep understanding of how LLMs fail.

That's why DeepCraft builds AI products differently from a generic development agency. GiantAI is our proof of work and everything we learned building it informs how we approach every AI integration and AI MVP we ship for clients.

Both repositories are open source. The platform app, the full AI agent codebase, is at github.com/ispolink/friendly-giant-ai-app. The smart contracts, the Base integration and on-chain mechanics are at github.com/ispolink/friendly-giant-contracts. The live platform is at giantai.ispolink.com.