Winner of the Founder Mode track at Agentic AI Build Week, one of the biggest agentic AI hackathons in the region, organized by GenAI Fund with AWS, VNG, and Tencent Cloud. Around 2,500 builders and 300+ teams pitched across the week in Ho Chi Minh City.

Syncnapse is an AI teammate that actually joins your meetings. It is not a recap tool that emails you bullets afterward. It is a real participant in Google Meet, Zoom, and Teams that you can talk to, that talks back, and that remembers. It captures not just what was said, but why decisions were made and who owns what, and turns all of it into governed company memory that is searchable, permissioned, and owned by the company instead of trapped in one person's head. When a task comes out of a meeting, it recommends whether that task should go to a human, an AI agent, or both, with actual reasoning attached. Built with my team, We Maid AI.

Under the hood, Syncnapse is a Turborepo monorepo (TypeScript, Node 22) running two distinct agent runtimes. The first is a chat-turn tool loop over the memory graph. The second is a real-time live meeting harness that takes raw audio through voice activity detection, speech-to-text, addressing and floor control, an LLM orchestrator, and finally text-to-speech, and it can run multiple bots at once with org-chart supervisor delegation. Recall.ai bots join the calls, and a Chrome extension (MV3, built with WXT) detects meeting tabs. A Hono and oRPC backend with Better Auth organizations persists everything to Postgres with pgvector (HNSW) through Drizzle, then ElectricSQL streams it back down into PGLite inside the React 19 TanStack Start dashboard, with Loro CRDT handling multi-agent merges. Governance is the core invariant: AI proposes, humans approve. Every memory write flows through a proposal state machine with versioning and audit trails, and the system integrates with Slack, Linear, and MCP through a save_to_synapse server. The whole stack is Dockerized and deployed on Railway.

EcoCraft AI (Rác Thái Xanh) is a Vietnamese-language web app that helps children turn recyclable materials into creative crafts. A child describes whatever recyclables they have in natural Vietnamese, the AI matches them against a verified craft library, and the app walks them through building it step by step with interactive 3D models. I built it for the Cuộc thi Sáng tạo thanh thiếu niên, nhi đồng (Youth Creativity Contest) in Đồng Nai, Vietnam.

There are two ways to tell the app what you have: a visual inventory grid of 12 material types with 3D previews, or a free-form Vietnamese chat where the AI pulls items out of natural language. A deterministic matching algorithm pairs materials to 8+ verified crafts, and a live AI coach (Thợ Cả AI) gives context-aware help while kids build. The stack is Next.js 16, React 19, Zustand, and Google model-viewer for 3D GLB rendering, with Groq (Llama 3 70B) as the AI and Ollama as an offline fallback. All 3D models are pre-generated with Meshy and served from Supabase Storage. One strict rule holds it all together: refuse, don't hallucinate. The AI never invents crafts that can't actually be built.

Architect is a structural health scanner and refactoring planner for vibe-coded projects. It scans a codebase across four dimensions, modularity (35%), security (25%), duplication (20%), and architecture (20%), and produces a quantified health score. It then installs YAML-based architectural blueprints into Claude Code, Cursor, Windsurf, or Copilot that teach best practices, separation rules, and anti-patterns for your exact tech stack.

It is written in TypeScript strict mode, using Babel Parser and tree-sitter for multi-language AST analysis, madge for dependency graph detection, and jscpd for code duplication scanning. The CLI ships 20+ stack-specific skills (Express, Next.js, Flask, FastAPI, Hono, and more) and supports phased refactoring workflows with approval gates, so AI agents can work through multi-step refactors without losing architectural coherence. It is published on npm as @levironexe/architect and comes with a Next.js 16 documentation site.

Credence is an AI agentic credit assessment system built for Swin Hackathon 2026. It goes after a real gap: 70 to 80% of Vietnamese micro-SMEs are denied credit because they lack formal financial data. At its core is a LangGraph-powered agent, an 18-node stateful graph that classifies loan officer intent, plans which tools to invoke, runs multi-step ML pipelines, and streams decisions back to the frontend in real time via SSE.

The agent orchestrates five specialized tools: an XGBoost credit scoring model (AUC-ROC 0.7705, tested on 46,127 profiles), a TreeSHAP explainer, a DiCE-ML counterfactual generator, a FairLearn fairness validator, and a data completeness checker. Unlike LLM-only systems, Credence uses true agentic control flow. It branches on intermediate results, re-assesses when data is incomplete, and combines ML outputs with RAG-based regulatory context before writing the final report. The result: it approves 2.3x more applicants than traditional models (42.5% vs 18.6%) while keeping the default rate lower (2.13% vs 2.22%).

Erudex is an AI-powered learning platform that turns natural language requests into structured, pedagogically ordered lesson plans. The interface is a split view: a docs panel on the left for navigating hierarchical lesson content (sections and subsections) and a chat panel on the right for requesting new lessons or refining existing ones. The AI always proposes a full lesson plan before anything touches the database, so you can accept or decline before committing.

It is built with Next.js 16 and the Vercel AI SDK, uses Claude Haiku for lesson generation, and enriches content live through Brave Search over the Model Context Protocol. Lesson plans, sections, items, and chat history are all persisted per user in Supabase PostgreSQL with Drizzle ORM. The whole project was developed spec-first with Spec Kit: every feature is designed, planned, and contracted in the specs/ directory before any implementation.

WriteFlow is a browser extension MCP client that connects to your Google Docs and gives you intelligent, context-aware document assistance. It helps you work with documents faster through semantic search, natural language queries, and automatic context understanding. Tab-aware intelligence means it always knows which document you are editing, answers come from RAG powered by a Redis vector database, and Google Drive integration is seamless. Production-scaled from day one, it is built for content writers, business professionals, researchers, and developers who want instant answers from their documents without the constant context switching.

The Itea Lab homepage is a modern, responsive website built with Next.js and Tailwind CSS to showcase the activities of the Itea Lab coding community. It highlights community projects, event listings, and member profiles, with server-side rendering for performance and SEO. It works as both an information hub and a community portal, a sleek and accessible place for developers to connect and collaborate.

SentinelAI is an AI-powered compliance automation framework that streamlines PCI DSS v4.0 audits for financial institutions, built during the VPBank Hackathon 2025. It brings together Amazon Bedrock AI (Claude 3.5 Sonnet), AWS Config, CloudTrail, Security Hub, and Aurora PostgreSQL to automate evidence collection, compliance evaluation, and report generation across 200+ AWS accounts.

Instead of manual auditing, teams get real-time dashboards, AI audit agents, cross-framework mapping, intelligent risk prioritization, and multi-format reporting. The architecture is modular: an AI agent orchestrator, an evidence collector, a RAG-based knowledge base, and a frontend compliance portal. With up to 90% faster evidence collection and 80% less prep time, SentinelAI shows what enterprise-scale, adaptive compliance management can look like.

GoBuy is a university project: a modern, multilingual marketplace built with Vue.js, Express.js, Supabase, and Tailwind CSS. It was designed for global accessibility from day one, with real-time language switching (English, Spanish, Vietnamese) and support for 10 currencies through live FastForex API integration.

The platform covers full e-commerce functionality: product listings with filtering, shopping cart management, checkout, and seller dashboards. Users sign in with email and password, magic links, or Google OAuth, and can customize their profiles.

A strong focus on modern UI/UX keeps the experience responsive and polished, with animations, loading skeletons, and instant translations. The backend REST API handles secure JWT authentication, Supabase-managed data, and smart caching for performance.

The frontend is deployed on Vercel and the backend on Render. Together they demonstrate a scalable architecture for international online marketplaces that balances accessibility, performance, and security.

Carbonio is a decentralized application for recording and verifying carbon emissions on the Solana blockchain, built for the BREAKOUT Hackathon by Colosseum. Companies register and upload their carbon data, third-party auditors verify it, and consumers can scan QR codes to see verified footprints. It is built with Next.js, TypeScript, Express, and Anchor smart contracts, uses IPFS (Pinata) and Supabase for storage, and supports NFT creation through Metaplex. With separate portals for consumers, companies, and auditors, it creates a multi-user ecosystem that tracks and rewards emission reduction.