SightShift AI

Inspiration

The world is filled with invisible changes — a crack in a bridge, a dent in a car, a rust patch on a machine — tiny shifts that go unnoticed until it’s too late. We wanted to build an AI that doesn’t just see but understands these transformations. SightShift AI was born from the idea of giving vision intelligence — the ability to compare, classify, and predict change, just like a human inspector, but infinitely faster and more reliable.

##What it does ##

SightShift AI is an advanced Visual Difference Intelligence Engine that detects, analyzes, and predicts changes in images or live video streams. It can scan cars, factories, infrastructure, or brand visuals — spotting defects, color inconsistencies, or rust formation — and generate spoken and written insights automatically. Users can simply point a camera or upload a photo sequence, and the system:

Detects differences using deep learning and image segmentation.

Classifies the type of change (scratch, crack, object shift, corrosion, logo mismatch, etc.).

Predicts future deterioration or deviation.

Generates AI-written inspection reports and voice summaries — no manual effort required.

Think of it as a digital inspector that never sleeps.

##How we built it ##

We designed SightShift AI as a hybrid system of real-time computer vision, predictive modeling, and natural language generation.

Tech Stack (Built With):

Frontend: React + Vite + Tailwind CSS + WebRTC (for live video capture)

Backend: FastAPI (Python) + RESTful API endpoints

AI/ML Core: TensorFlow + PyTorch + OpenCV + scikit-learn

Vision Models: YOLOv8 for live defect detection, SSIM & image differencing for subtle changes

Prediction Engine: Time-series forecasting (LSTM) for deterioration trends

AI Assistant: OpenAI + Whisper + ElevenLabs for report narration and summaries

Database & Storage: Firebase / Supabase + Google Vision API integration

Visualization: Streamlit + Chart.js + Heatmap.js

Reporting: jsPDF + ReportLab for instant PDF reports

Deployment: Vercel (Frontend) + Render / AWS Lambda (Backend)

Edge Support: ONNX Runtime for local inference when offline

Every module was optimized for accuracy, speed, and low-latency inference, enabling instant detection without specialized hardware.

Challenges we ran into

Precision under chaos: Getting consistent defect detection despite lighting, angle, and texture differences.

Model interpretability: Teaching the AI why a visual change matters — not just that it exists.

Real-time voice feedback: Syncing AI-generated summaries with live camera inference.

Scalability: Ensuring smooth operation across multiple devices without GPU dependence.

Privacy: Designing encrypted pipelines for sensitive industrial or brand imagery.

Accomplishments that we’re proud of

Built a general-purpose visual change engine that works across industries — not tied to any single domain.

Achieved 90%+ accuracy in detecting fine defects and visual inconsistencies.

Created a voice-enabled AI inspector that can narrate its findings live.

Delivered end-to-end results (visual + textual + auditory) without any hardware integration.

Designed a future-ready architecture that can scale to drones, CCTV, or AR glasses.

What we learned

We learned that true innovation lies in observation. Seeing the “before and after” isn’t enough — AI must interpret meaning. We explored how time-series AI can extend vision beyond the present — predicting future changes from past data. We also discovered the importance of explainability — users trust AI more when it can explain what it sees in plain language.