Real-Time Face Detection (Edge Deployment)
Hybrid CNN + Haar for low-light robustness on FERET dataset.
What I Did:
- Implemented ensemble model achieving 95% accuracy on diverse lighting conditions.
- Optimized inference on Jetson Nano hardware, reducing latency by 3× via ONNX export.
- Deployed as Streamlit app for real-time security monitoring.
Demo: Upload image below—watch detection in action.
Live Demo Embed Here (e.g., Gradio/Streamlit iframe)
PyTorch, OpenCV, ONNX
Multilingual Voice Assistant
PyTorch STT with fine-tuned Whisper for 3 languages.
What I Did:
- Fine-tuned Whisper model on custom multilingual dataset, boosting speed by 40%.
- Integrated speech-to-text with intent recognition for query handling.
- Hosted on Hugging Face Spaces for interactive testing.
Demo: Speak into mic—get responses in Hindi/English/Tamil.
Live Voice Demo Embed Here
PyTorch, Whisper, Gradio
Fraud Anomaly Detector
Isolation Forest + XGBoost on transaction data, API-served.
What I Did:
- Built end-to-end pipeline with 98% precision on imbalanced datasets.
- Deployed as FastAPI endpoint for real-time scoring.
- Handled 10K+ txns/min with monitoring dashboard.
Demo: Input txn data—see anomaly flags.
Live API Demo Embed Here
Scikit-Learn, XGBoost, FastAPI
BERT Sentiment Pipeline
Transformers for social monitoring, freelance-validated.
What I Did:
- Fine-tuned BERT for 92% F1 on tweet sentiment, processing 1K+/min.
- Added RAG for context-aware analysis.
- Integrated Gradio UI for client feedback loops.
Demo: Paste tweet—get sentiment score.
Live Sentiment Demo Embed Here
Hugging Face, BERT, Gradio
Crop Yield Forecaster
LSTM on satellite time-series for climate-adaptive agriculture.
What I Did:
- Developed LSTM model hitting 85% accuracy on NASA satellite data.
- Incorporated climate variables for resilient predictions.
- Created Binder notebook for interactive forecasting.
Demo: Select region/date—view yield forecast.
Interactive Notebook Embed Here
TensorFlow, LSTM, Pandas