Deepank Dixit

With a solid AI foundation from the Indian Institute of Science, complemented by practical Security and AI engineering at Cisco, I enhance cybersecurity using advanced machine learning. My expertise extends from generative modeling with Transformers, GANs and VAEs to the deployment of NLP-driven conversational AI for optimizing Cisco's security infrastructure. Technical writer and speaker at premier industry events like the Cisco Data Science Summit, Offensive Security Summit, and QCon Plus.

1. Syslog Anomaly Detector: VAE-Enhanced Security Analytics for Cisco ISE [link]

Pioneered the use of semi-supervised Variational Autoencoders (VAEs) to pinpoint and examine anomalies within Cisco ISE syslog data, distinguishing irregular network behaviors from standard operations. Achieved a more efficient log analysis process by isolating 0.3% of data sequences representing significant anomalies, leading to a targeted threat detection strategy and the conservation of SME man-hours.

2. Intelligent Document Assistant: MultiPDF-ChatMaster [link]

Built an LLM and RAG based chatbot for handling multiple PDFs for content extraction and interactive assistance, using LangChain for text splitting and OpenAI for embeddings. Adopted FAISS indexing for efficient similarity searches in large-scale documents, providing precise answers from PDF text chunks, enhancing user query resolution and information retrieval.

3. Web Interaction Navigator: LangStream-WebBot with RAG [link]

Developed a web-interfacing chatbot employing RAG and GPT-4’s language understanding through LangChain. Improved user experience with a Streamlit-based GUI that combines conversational AI with real-time data extraction and interaction, supported by ChromeDB for robust vector storage and quick data retrieval.

4. Neural Image Caption Generator [link]

Leveraged Transfer Learning with the Inceptionv3 CNN to encode images from the Flickr8k Dataset, serving as the foundation for an Encoder. Integrated processed text captions into a TransformerDecoder equipped with six Multi-Head Attention Layers, utilizing Keras for word embeddings. The model, tuned with Sparse categorical cross-entropy loss and Adam optimizer, achieved around 50% test accuracy in generating descriptive captions.

5. Countering Adversarial Attacks in ML [link]

Utilized PyTorch to develop a CNN-based MNIST classifier, then exposed its vulnerabilities using adversarial techniques such as the Fast Gradient Sign Method (FGSM) and One-pixel attacks, which initially reduced test accuracy from 98.59% to 57.96%. Strengthened the model's defenses through adversarial training, restoring accuracy on adversarial examples to 98%. Demonstrated the effectiveness of these strategies through detailed visualizations.

6. Bayesian ML for Environmental Data Recovery [link]

Devised Bayesian ML models to infer missing environmental data across a network of 5 sensor nodes. Applied Bayesian linear regression to predict distributions for temperature and humidity, enhancing accuracy with spatial-aware methods like Kriging. Outperformed standard models like Linear Regression model and Random Forest Regressor, in precision, while addressing computational and integration complexities inherent to Bayesian Neural Networks and Gaussian processes.

7. Exoplanet Classification & Analysis Suite for Habitability Insights [link]

Implemented probabilistic machine learning models to interpret exoplanetary data, employing Bayesian Regression and Chow-Liu algorithms for intricate parameter relationships and habitability assessment. Created a classification system to effectively size exoplanets, surpassing baseline accuracy with a 47% success rate. Utilized Gaussian Mixture Models and PCA for clear visualization of habitable zones, overcoming incomplete data with advanced imputation techniques like GaussianProcessRegressor.

8. Time Series Weather Forecasting [link]

Utilized seven years of climate data from the 'jena_climate_2009_2016.csv' dataset to forecast future temperatures. Employed and benchmarked various recurrent neural network models, including RNN, LSTM, and GRU, against a non-ML baseline, effectively comparing their predictive capabilities.

DeepLearning.AI Certifications

• Building and Evaluating Advanced RAG
• Finetuning Large Language Models
• LangChain Chat with Your Data
• Generative AI with Large Language Models

Cisco Certifications

• Cisco Certified DevNet Associate
• Cisco Certified Network Associate Routing and Switching (CCNA)
• Cisco Certified Network Professional Security (CCNP Security)
• Cisco Certified Specialist - Security Identity Management Implementation (Cisco ISE)

• Speaker at Cisco Data Science Summit 2024: "GANs for Realistic Cyber-Attack Simulations."
• Speaker at Cisco Offensive Summit on "LLMs Unbound: Navigating the Data Privacy Conundrum" (2023), "Understanding Security for Internet of Behavior" (2022), and "Trust Me, I'm an Insider – Zero Trust Security" (2021).
• Speaker at Cisco SecCon 2021 and QCon Plus 2021 on Zero Trust Security.
• Awarded the Data Science Green Belt at Cisco; first place and runner-up in several ‘Capture The Flag (CTF)’ challenges.
• Recognized as a Cisco Security Interstellar Pioneer (Security Ninja Black Belt) for outstanding contributions to security.