Projects

Temporal deepfake detection with a ResNet18 + LSTM model trained on 3,200+ videos.

• Loaded 3,200+ raw videos from AWS S3 and extracted and preprocessed frames with OpenCV for supervised training.
• Trained a ResNet18 CNN + LSTM model for temporal deepfake detection: 84% precision and 84% recall.
• Batched the PyTorch inference pipeline to reach under 500 ms per frame for near-real-time analysis.

Real-time intrusion detection over 1,000+ packets per second.

• Captured 1,000+ packets per second of live network traffic with Scapy and extracted features for real-time classification.
• Trained an XGBoost model to detect malicious traffic: 92% precision and 87% recall in live evaluation.
• Deployed the inference pipeline on streaming data for low-latency monitoring and alerting.

A natural-language-to-SQL agent over PostgreSQL with a local LLM.

• Built a natural-language-to-SQL pipeline with LangChain and LangGraph for conversational querying of PostgreSQL.
• Reached under 200 ms query times on 3,000+ row tables using psycopg2 and Supabase.
• Used Mistral-7B via Ollama to turn query results into short, business-relevant summaries.

A logistic regression desktop app that classifies movement from CSV data.

• Built a desktop app for movement classification using logistic regression.
• Added a Tkinter GUI for CSV import, feature extraction and labeled output.
• Reached over 97% accuracy.