Smart Lab — Building an AI-Powered Academic Research and Growth Platform

Inspiration

As STEM students, we face an overwhelming amount of learning materials every day: from hundreds of pages of lecture slides to technical papers, from complex mathematical derivations to experimental code. Traditional learning methods are inefficient, and knowledge becomes fragmented. We asked ourselves: Can we use AI to fundamentally redefine "learning" itself?

This inspiration came from our real pain points:

📚 Slide Overload: A single course may have 200+ pages of slides; manually organizing notes is time-consuming and prone to missing key points
📖 Difficulty Reading Papers: Top conference papers are filled with jargon, lacking systematic interpretation tools
🔍 Inefficient Literature Review: When writing our thesis, just "finding literature" consumed 40% of our time
🤝 Resource Silos: Students lack efficient mechanisms for sharing and collaborating on research resources
📋 Lost Progress: Learning paths are unclear—"What should I learn after finishing this?"

Smart Lab was born to leverage the power of the Gemini large language model to solve these real learning challenges.

What It Does

Smart Lab is a one-stop AI-driven academic growth platform covering the entire workflow from research reading to knowledge management, from resource exchange to learning planning.

🔬 Research Suite

Module	Features	Problem Solved
Smart Reading	Multi-source comparative reading, page-by-page Q&A, automatic analysis report generation	Compress paper reading time from 3 hours to 30 minutes
Paper Agent	Full-text structured decomposition, core contribution extraction, methodology visualization	Quickly understand the core innovations of complex papers
Literature Agent	Intelligent literature search + automatic review generation	One-click draft of "Related Work" section
Research Agent	End-to-end generation from idea to technical tutorial	Quickly transform ideas into actionable learning paths

📖 Learning Suite

Module	Features	Problem Solved
Smart Disassembly	Intelligent PDF segmentation → Knowledge module extraction → Exercise correlation → Pre-exam quiz generation	Complete structured learning of 100-page slides in 10 minutes
Study Companion	Page-by-page PDF Q&A, personalized explanations, note accumulation	Like having a 24/7 private tutor
Note Center	Bidirectional linking, knowledge graph visualization, tag system	Build personal knowledge base, achieve "Second Brain"
Course Architect	One-click generation from topic to course outline	Quickly create systematic self-study materials

🌐 Smart Square (Resource Exchange)

Feature	Description
Research Ops	Professors post research internships → Students submit applications → One-click resume generation
Corporate Ops	Companies post industry project needs → Students form teams to respond
Resource Ops	Crowdsourced sharing of quality courses/notes/tutorials → Points incentive mechanism
Channels	Project-based real-time group chat collaboration → Slack-like communication experience
Consultation	Book 1-on-1 consultations with industry experts or Open Day live streams

📋 Task Management

Feature	Description
Learning Center	Skill tree visualization → Course progress tracking → Task list management
Home Dashboard	At-a-glance view: recent activities, statistics, quick access
SkillTree Component	Interactive skill learning path map based on React Flow

How We Built It — Powered by the Google Ecosystem

This project was built entirely within the Google ecosystem, leveraging a powerful trio of tools that seamlessly integrate across the entire development lifecycle:

┌─────────────────────────────────────────────────────────────────────────────┐
│                     THE GOOGLE ECOSYSTEM ADVANTAGE                           │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                              │
│   🚀 BUILD           │   🔧 REFINE           │   ☁️ DEPLOY                   │
│   ────────────────   │   ────────────────    │   ────────────────            │
│   Google AI Studio   │   Antigravity (Gemini │   Google Cloud Run            │
│   + Gemini API       │   Code Assist)        │   + Cloud Build               │
│                      │                       │                               │
│   • Initial proto-   │   • Intelligent code  │   • Containerized             │
│     typing in AI     │     refactoring       │     deployment                │
│     Studio           │   • Feature expansion │   • Auto-scaling              │
│   • Export to code   │   • Bug fixing with   │   • Global CDN                │
│   • Gemini 3 Flash   │     AI assistance     │   • CI/CD pipeline            │
│     as core engine   │   • Code quality      │   • Secret management         │
│                      │     improvements      │                               │
└─────────────────────────────────────────────────────────────────────────────┘
              │                     │                     │
              └─────────────────────┼─────────────────────┘
                                    │
                              UNIFIED GOOGLE
                              DEVELOPER EXPERIENCE

🚀 Phase 1: Rapid Prototyping with Google AI Studio

The journey began in Google AI Studio, where the initial concept was rapidly prototyped:

Interactive Prompt Engineering: Tested and refined prompts for Smart Reading, Paper Agent, and Literature Agent directly in the AI Studio interface
One-Click Code Export: The initial React application was exported directly from AI Studio, providing a production-ready foundation
Gemini API Integration: Seamlessly integrated @google/genai SDK with the latest Gemini models (gemini-3-flash-preview, gemini-2.0-flash)

// Direct integration with Gemini API from exported AI Studio code
import { GoogleGenAI } from "@google/genai";
const client = new GoogleGenAI({ apiKey: process.env.API_KEY });

🔧 Phase 2: Intelligent Development with Antigravity

Once the foundation was in place, Antigravity (the Gemini-powered code assistant) became the primary development partner:

Feature Expansion: Added 7 complete AI agent modules with Antigravity's intelligent code generation
Complex Refactoring: Implemented Map-Reduce architecture for large PDF processing with AI-assisted design
Bug Detection & Fixing: Identified and resolved edge cases in streaming response handling
Bilingual Support: Implemented internationalization with batch translation using batchTranslateTexts
Code Quality: Maintained 100% TypeScript type safety with 492+ lines of strongly-typed interfaces

"Antigravity wasn't just a code assistant—it was a pair programming partner that understood the entire codebase context and could implement complex features across multiple files simultaneously."

☁️ Phase 3: Production Deployment with Google Cloud

The final step was deploying to Google Cloud Run, completing the end-to-end Google ecosystem experience:

Cloud Build: Automated CI/CD pipeline with cloudbuild.yaml configuration
Container Registry: Docker images stored in Google Container Registry
Cloud Run: Serverless deployment with auto-scaling and global edge caching
Secret Management: API keys handled securely via build-time substitutions

# cloudbuild.yaml - Seamless deployment pipeline
steps:
  - name: 'gcr.io/cloud-builders/docker'
    args: ['build', '--build-arg', 'GEMINI_API_KEY=${_GEMINI_API_KEY}', '-t', 'gcr.io/$PROJECT_ID/smart-lab:latest', '.']
  - name: 'gcr.io/google.com/cloudsdktool/cloud-sdk'
    args: ['run', 'deploy', 'smart-lab', '--image', 'gcr.io/$PROJECT_ID/smart-lab:latest', '--allow-unauthenticated']

Deployment URL: https://smart-lab-3jid3va6la-de.a.run.app

Why Google Ecosystem Matters

Benefit	Description
Seamless Integration	From prototyping to deployment, all tools speak the same language
Unified Authentication	Single Google account manages AI Studio, Cloud Console, and Antigravity
Consistent AI Models	Same Gemini models power the IDE assistant and the production application
Cost Efficiency	Cloud Run's pay-per-use model + free tier = minimal hosting costs for startups
Developer Velocity	AI-assisted development reduced feature implementation time by 5x

Technology Stack

Frontend:  React 18 + TypeScript + Vite
UI:        TailwindCSS + Custom Design System
AI Core:   Google Gemini API (gemini-3-flash-preview / gemini-2.0-flash)
PDF:       pdf-lib (PDF processing) + html2pdf.js (export)
Storage:   IndexedDB (local persistence)
Markdown:  react-markdown + remark-gfm
Charts:    React Flow (skill tree)
Deploy:    Google Cloud Run + Cloud Build + Container Registry

Core Architecture

┌─────────────────────────────────────────────────────────┐
│                    App.tsx (Routing Layer)              │
├─────────────────────────────────────────────────────────┤
│   GlobalSidebar   │        Content Area                 │
│   ───────────────────────────────────────────────────── │
│                   │  Home / LearningCenter / SmartSquare │
│                   │  SmartReading / SmartDisassembly     │
│                   │  PaperAgent / LiteratureAgent        │
│                   │  CourseArchitect / ResearchAgent     │
│                   │  NoteCenter / StudyCompanion         │
└─────────────────────────────────────────────────────────┘
          │                       │
          ▼                       ▼
    ┌─────────────────────────────────────────────┐
    │              Services Layer                 │
    ├─────────────────────────────────────────────┤
    │  geminiService.ts    (AI call wrapper)      │
    │  storageService.ts   (IndexedDB CRUD)       │
    │  paperService.ts     (Paper parsing logic)  │
    │  disassemblyService.ts (Slide disassembly)  │
    │  agentService.ts     (Multi-step generation)│
    └─────────────────────────────────────────────┘
          │
          ▼
    ┌─────────────────────────────────────────────┐
    │           Google Cloud Infrastructure       │
    ├─────────────────────────────────────────────┤
    │  Cloud Run       (Serverless hosting)       │
    │  Cloud Build     (CI/CD pipeline)           │
    │  Container Reg   (Docker image storage)     │
    │  Gemini API      (AI inference endpoint)    │
    └─────────────────────────────────────────────┘

Key Design Decisions

Map-Reduce Architecture (SmartDisassembly): Split PDFs of hundreds of pages into modules, process in parallel, then merge—significantly improving generation speed
Streaming Output: All AI conversations use generateContentStream, so users don't have to wait for complete responses
Local-First Storage: Using IndexedDB to save session records—works offline and protects privacy
Bilingual Support: Chinese-English switching via UI_TRANSLATIONS, batch translation using batchTranslateTexts
Serverless Deployment: Cloud Run + Cloud Build enables zero-ops deployment with automatic scaling

Challenges We Ran Into

Challenge 1: Token Limits for Large PDF Processing

Problem: A single PDF may contain 100+ pages; sending it directly to Gemini would exceed the token limit.

Solution:

Design a DisassemblyPlan planning phase where AI first plans "which pages belong to which module"
Then process in batches by module granularity, with each module containing only 10-20 pages of physical slices
Finally merge to generate the complete knowledge notebook

// services/disassemblyService.ts
interface DisassemblyPlan {
  course_title: string;
  modules: {
    id: string;
    title: string;
    page_start: number;
    page_end: number;  // ← Physical slice boundary
    focus_points: string[];
  }[];
}

Challenge 2: Citation Accuracy in Literature Reviews

Problem: AI-generated literature reviews may contain "hallucinated references."

Solution:

Adopt Grounded Generation: AI must generate content based only on actual search results
Force generateLiteratureReview to use searchReferences as the sole information source
Output format enforces Citation Markers [1], [2]... for easy manual verification

Challenge 3: Identity Verification for Smart Square

Problem: Need to distinguish permissions between "student submitting resume" and "professor reviewing applications."

Solution:

Define UserRole: 'student' | 'researcher' | 'company' | 'admin'
Dynamically render UI buttons based on role (e.g., only post authors can see "View Applicants" button)
Use Application state machine to manage the application workflow: PENDING → APPROVED/REJECTED

Challenge 4: Persisting Complex State

Problem: SmartDisassembly sessions need to save: PDF files, planning data, module content, quiz data, and other multi-layered nested states.

Solution:

Design a flattened DisassemblySession type, storing PDFs as Base64 in IndexedDB
Use optimistic updates (updateSession) to ensure UI responsiveness
Add status field to support resumable sessions: 'uploading' | 'planning' | 'generating' | 'completed'

Accomplishments That We're Proud Of

✅ Complete AI-Assisted Research Workflow: From literature search → paper reading → note management → review generation, forming a closed loop

✅ Map-Reduce Architecture for Smart Disassembly: Innovatively transforms long document processing into parallelizable modular tasks

✅ Fully Functional Community Features: Not just a simple post list, but a complete collaboration system including application workflows, resume uploads, and project channels

✅ 7 Core Agent Modules: Each module has independent Prompt Engineering and business logic

✅ 100% TypeScript Type Safety: 492 lines of types.ts ensure consistency of data structures

What We Learned

Technical Level

Prompt Engineering is a Core Competency: The same functionality with better prompts can improve output quality by 3x
User Experience Value of Streaming Output: Nobody wants to stare at a loading screen for 30 seconds
"Small Brain" Strategy for Large Models: Complex tasks should be decomposed into multiple small tasks rather than having AI complete everything at once

Product Level

Quality Over Quantity: Seven modules seem independent but are seamlessly connected through onNavigate and Deep Links
Pain-Point-Driven Development: Every feature stems from real learning scenarios, not "technical showoff"

Collaboration Level

AI is a Co-Pilot, Not an Autopilot: The platform's positioning is to enhance human capabilities, not replace thinking

What's Next for Smart Lab

Phase	Plan
Short-term	Integrate more Gemini models (e.g., Gemini Ultra) to enhance complex reasoning capabilities
Mid-term	Add cloud sync functionality for cross-device data consistency
Long-term	Build an academic social graph for intelligent collaborator recommendations based on research interests
Vision	Become the one-stop growth partner for every researcher from "beginner → advanced → output"