HackStack

Project Info

assembl3D

Devpost

Inspiration

We've all been there, sitting on the floor surrounded by hundreds of screws, mysterious wooden slabs, and an incomprehensible IKEA manual that looks like it was designed by aliens. Assembly manuals are notoriously difficult to follow: tiny diagrams, confusing arrows, and no way to see what the final step should actually look like in 3D space. We wanted to turn this frustrating experience into something intuitive and interactive. What if you could search for any furniture product, automatically get its assembly manual, and see each step visualized in an interactive 3D environment? That's the vision behind assembl3D - your AI-powered copilot for furniture assembly.

What it does

assembl3D is an end-to-end platform that makes furniture assembly effortless: Search & Discover: Search for any furniture product through browsing the 50+ most popular IKEA products, or even paste any product URL. Intelligent Scraping: Our system uses Bright Data's powerful APIs to populate our library for Ikea products and automatically search Google for products (SERP API), scrape product pages to find important metadata, download protected PDFs (Web Unlocker), and collect product images - all without manual work. AI-Powered Processing: Google Gemini AI, analyzes each page of the PDF manual and extracts step-by-step assembly instructions with clear descriptions, required parts with quantities and dimensions, necessary tools, 3D positioning data, and assembly actions with animations. Interactive 3D Visualization: Beautiful 3D viewer displays each assembly step with real-time rendering using React Three Fiber, intuitive step-by-step navigation, visual parts lists, required tools, smooth animations showing how parts fit together, and orbit controls to view from any angle. AI Assembly Assistant: Reka AI-powered chatbot answers questions about the current step, helps identify parts and tools, provides troubleshooting assistance, and gives contextual advice in real-time. As well as helps for cross-checking assembly extraction and spatial positioning of parts.

How we built it

We built a sophisticated PDF processing pipeline that renders pages of 2D drawings in 3D, extracts them as optimized images, and uses MD5 hashing to prevent duplicate processing. Our AI vision system uses carefully crafted prompts to guide Gemini in extracting structured JSON from complex assembly diagrams. For 3D rendering, we generate geometric primitives procedurally from AI-extracted dimensions rather than using pre-made models, enabling us to render any part type on the fly. Our web scraping strategy uses SERP API for product information, with smart rate limiting and caching to optimize costs.

Challenges we ran into

PDF Complexity: Assembly manuals are primarily visual with complex diagrams. We pivoted from text extraction to converting pages to high-resolution images for AI vision analysis, which worked significantly better. Rate Limiting & Costs: Both Gemini (60 requests/minute) and Bright Data (pay-per-request) have limits. We implemented 500ms delays between requests, MD5-based caching to avoid reprocessing, and smart scraping that only downloads new products. Coordinate Systems: Converting PDF positions to Three.js 3D coordinates was complex due to different coordinate systems (Y-up vs Z-up). 3D Performance: Rendering complex assemblies with 50+ parts was initially slow. We optimized using low-poly primitives, frustum culling, lazy loading, and shader optimization to achieve smooth performance.

Accomplishments we're proud of

AI Vision Breakthrough: Successfully getting Gemini to understand complex IKEA diagrams and extract structured data (parts, quantities, sequences, tools) was a major achievement. This opens possibilities for processing any visual instruction manual. Beautiful, Professional UI: Our frontend is polished with smooth animations, fully responsive design, interactive 3D controls, and visual feedback throughout. It looks like a production app, not a hackathon project. Real Product Library: We scraped and cached 50 real IKEA products with actual images, automatically categorized by room type, providing immediate value to users without requiring searches.

What we learned

Technical Skills: We learned how powerful modern vision AI like Gemini is at understanding complex diagrams with proper prompting. Bright Data's APIs taught us professional web scraping - SERP API abstracts search result parsing, Web Unlocker handles proxies and CAPTCHAs automatically, and proper rate limiting with caching is essential for cost management. We deepened our understanding of Three.js, React Three Fiber, geometric primitives, and real-time 3D performance optimization. TypeScript's strong typing across the full stack prevented countless bugs and made refactoring under time pressure much easier. Product & Design: We focused on solving a real, universal problem (confusing manuals) rather than showcasing technology. This user-first mindset guided all decisions. We implemented progressive enhancement so the app works with cached data even when APIs are unavailable. We learned that small details like loading states, smooth transitions, and hover effects dramatically improve perceived quality and professionalism.

What's next

Short-term (3 months): Improve position parsing to extract actual 3D coordinates from diagrams. Add advanced animation system with play/pause controls and sequential part movements. Expand to support multiple furniture brands beyond IKEA. Implement mobile AR integration using WebXR to overlay instructions on real furniture. Add user accounts with progress tracking, notes, and sharing.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only
  • CSSIn code
  • ExpressIn code
  • Google GeminiIn code
  • JavaScriptIn code
  • Next.jsIn code
  • ReactIn code
  • Tailwind CSSIn code
  • TypeScriptIn code
  • Node.jsClaimed
  • PythonClaimed

8 of 10 appear in the indexed code. 2 claimed on Devpost could not be matched to code, which may simply mean the tool leaves no trace in the repository.

AI coding agents

No AI coding agent signals were found in this repository.

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

322 KB

Source files

90

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.

0 stars