README.md

metadata

title: GharScan
emoji: 🏗️
colorFrom: gray
colorTo: red
sdk: gradio
sdk_version: 6.14.0
app_file: app.py
pinned: true
license: mit
short_description: AI Building Defect Inspector for India
tags:
  - computer-vision
  - building-inspection
  - defect-detection
  - india
  - minicpm-v
  - openbmb
  - gradio
  - backyard-ai
  - build-small-hackathon
  - track:backyard
  - sponsor:modal
  - achievement:offgrid
  - achievement:welltuned
  - achievement:offbrand
  - achievement:llama
  - achievement:sharing
  - achievement:fieldnotes
models:
  - ritvik360/gharscan-qwen2vl-lora
  - ritvik360/gharscan-qwen2vl-gguf
datasets:
  - ritvik360/gharscan-defect-dataset
  - ritvik360/gharscan-agent-traces

🏗️ GharScan — AI Building Defect Inspector for India

Track: Backyard AI · Build Small Hackathon 2026

"My neighbor Aunty Puja, a retired teacher in a 1982 DDA flat in Delhi, had three masons give her three different quotes for a crack she didn't understand. GharScan told her in 12 seconds: settlement crack, Severity 2/5, not structural, seal before monsoon — ₹400–600, local mason."

---

What It Does

Point your phone camera at any defect in your home — a crack, a damp patch, rust stains, salt deposits — and get an instant expert-grade triage report:

• Defect type from an 8-class Indian residential taxonomy
• Severity score (1–5, color-coded) with structural risk flag
• Immediate action in plain language (English or Hindi)
• Cost estimate in INR from a 2026 Delhi/NCR market rate matrix
• Who to call: painter / mason / waterproofing contractor / civil engineer
• Monsoon risk flag (critical for Indian users pre-June)

The Problem

Over 62% of India's urban housing was built before 1990. When a homeowner sees a wall crack, they have three options: pay ₹2,000–8,000 for a civil engineer, ask a mason who has a conflict of interest, or Google it and get scared. GharScan is the affordable first answer.

Technical Architecture

Layer	Technology
Base model	Qwen2-VL-2B-Instruct (2.07B params)
Fine-tuning	LoRA r=16, 7 modules, Modal A100-80GB, 3 epochs
Dataset	8,860 deduplicated images → 17,720 VQA records
Deduplication	CLIP ViT-B/32, cosine sim >0.95 threshold
Inference	HuggingFace ZeroGPU · gr.Blocks · no external API calls
Cost estimation	Deterministic INR lookup table (not model-generated)
Offline runtime	llama.cpp GGUF Q4_K_M · 941MB
Agent tracing	Auto-uploaded to ritvik360/gharscan-agent-traces

How to Use on Mobile

1. Open this Space on your phone browser
2. Tap Take Photo — your rear camera opens directly
3. Photograph the defect (crack, stain, damage)
4. Wait ~12 seconds for analysis
5. Read your inspection report

REQ Compliance Checklist

Requirement	Status	Evidence
REQ-01: ≤32B parameters	✅	Qwen2-VL-2B = 2.07B params
REQ-02: Gradio Space in org	✅	GharScan Space
REQ-03: Demo video	✅	Demo Walkthrough
REQ-04: Social media post	✅	Social Post
REQ-05: ZeroGPU limit	✅	1 ZeroGPU Space used
REQ-06: README tags	✅	YAML above includes all tracks + badges

Bonus Quest Badges Claimed

Badge	Status	Proof
🔌 Off the Grid — no cloud APIs	✅	ZeroGPU only, inference.py has zero external API calls
🎯 Well-Tuned — published fine-tune	✅	ritvik360/gharscan-qwen2vl-lora
🎨 Off-Brand — custom UI beyond defaults	✅	Dark concrete-grey theme, custom CSS, HTML report cards
🦙 Llama Champion — GGUF + llama.cpp	✅	ritvik360/gharscan-qwen2vl-gguf · Q4_K_M 941MB
📡 Sharing is Caring — agent traces	✅	ritvik360/gharscan-agent-traces
📓 Field Notes — blog post	✅	Read on HuggingFace Blog

All 6 badges claimed → Bonus Quest Champion eligible

Repositories

Resource	Link
🤖 Fine-tuned LoRA	ritvik360/gharscan-qwen2vl-lora
📦 GGUF (Q4_K_M, 941MB)	ritvik360/gharscan-qwen2vl-gguf
📊 Training dataset	ritvik360/gharscan-defect-dataset
🔍 Agent traces	ritvik360/gharscan-agent-traces
📝 Blog post	huggingface.co/blog/ritvik360/gharscan

Citation

If you use the GharScan dataset or model:

@misc{gharscan2026,
  title  = {GharScan: AI Building Defect Inspector for Indian Residential Properties},
  author = {ritvik360},
  year   = {2026},
  url    = {https://huggingface.co/spaces/build-small-hackathon/gharscan}
}