Deepfake Detection using CNN-LSTM
Technologies: PyTorch, OpenCV, ResNet-18, Bi-LSTM
Timeline: 2025

Objective: Develop a robust deepfake detection system using hybrid deep learning architecture.

Approach: Built a hybrid deep learning model combining ResNet-18 CNN for spatial feature extraction (512-D embeddings) and Bidirectional LSTM (256 hidden units, 2 layers) for temporal pattern analysis across 16-frame video sequences. Implemented comprehensive face preprocessing pipeline using dlib HOG detector with 68-point landmark alignment, frame sampling at 1 FPS, and augmentation strategies (geometric rotation ±15°, photometric brightness ±20%).

Results: Achieved 87.69% accuracy and AUC-ROC ≈ 0.90 on 2,502 sequences (1,308 real, 1,194 fake). Trained with AdamW optimizer, cross-entropy loss with class weighting, and ReduceLROnPlateau scheduler. Achieved robust balanced generalization with F1: 88.28% (Real), 87.04% (Fake).

Key Contributions: Modular architecture for feature extraction and temporal analysis; comprehensive evaluation methodology; reproducible training pipeline with detailed hyperparameter documentation.

Automated Lunar Traverse Planner
Technologies: Python, NumPy, Matplotlib, YOLO crater detection
Timeline: 2025 | LAM Challenge 2025 Finalist (Top 0.3%)

Objective: Design an intelligent path planning system for autonomous lunar rover missions.

Approach: Developed a multi-layered system integrating YOLO-based crater detection with solar illumination-aware navigation. Implemented physics-based solar shadow simulation using sun elevation and azimuth angles for real-time illumination assessment. Designed three-tier hierarchical navigation: Tier 1 (science-optimal direct path), Tier 2 (rotational escape search in 8 directions), Tier 3 (last-resort hazard breach protocol with relaxed safety margins). Automated scientific waypoint selection using crater rim geology scoring and regolith sampling on 15m grid with ≥20m crater exclusion radius.

Results: Successfully navigated 100m² terrain maps with optimal solar efficiency (≥0.90 threshold) maintained throughout traversal. System demonstrated robust adaptation to complex terrain scenarios and dynamic constraint satisfaction.

Recognition: Selected as Finalist (Top 0.3%, 24 out of ~8,000 registrations) in LAM Research Challenge 2025.

BDH Narrative Consistency Evaluator
Technologies: PyTorch, Transformers, Mechanistic Interpretability
Timeline: 2025

Objective: Develop interpretable transformer models for narrative consistency detection with mechanistic analysis.

Approach: Built complete inference and interpretability pipeline on Baby Dragon Hatchling (BDH) transformer architecture (6-layer, d=512, 8 heads, RoPE positional embeddings). Engineered structured training data via synthetic contradiction generation using safe name-swapping, template-based input formatting, and balanced positive/negative sampling from literary novels. Implemented mechanistic interpretability probes: captured per-layer, per-head attention scores; averaged across heads; generated high-resolution attention heatmaps to visualize model decision pathways across layers and heads.

Pipeline: End-to-end workflow from pretraining on raw novel text through supervised fine-tuning to inference with GPT2TokenizerFast. Classifier uses max-pooled hidden states with MLP head for final predictions.

Key Contributions: Mechanistic interpretability probes for transformer analysis; systematic approach to understanding model decision boundaries; reproducible interpretation methodology.

Image Segmentation using U-Net
Technologies: PyTorch, Albumentations, Semantic Segmentation
Timeline: 2025

Objective: Implement state-of-the-art semantic segmentation pipeline with modern best practices.

Approach: Implemented complete semantic segmentation pipeline with U-Net architecture incorporating custom DataLoader with train/validation splits and real-time augmentations via Albumentations. Applied Dice + BCE hybrid loss function for balanced optimization. Integrated GPU acceleration with AMP (Automatic Mixed Precision), reproducible seeding for deterministic results, and modular architecture for scalable experimentation.

Results: Achieved high Dice coefficient and Intersection-over-Union (IoU) scores with comprehensive visualization of prediction masks and ground truth comparisons.

Key Features: Production-ready data pipeline; modern loss functions; GPU-optimized training; visualization utilities for result analysis and interpretation.

Neural Style Transfer via VGG19
Technologies: PyTorch, OpenCV, VGG19, Anime Art Style
Timeline: 2024

Objective: Implement neural style transfer to convert images to anime art style.

Approach: Implemented neural style transfer with pretrained VGG19 backbone, separating content and style using multi-layer activations. Minimized combined Content + Style + Total Variation (TV) loss through gradient-based optimization. Tuned style/content weights, learning rate, and image resolution for optimal aesthetic results. Enabled GPU acceleration with AMP support and reproducible random seeds for consistent output.

Results: Successfully generated high-quality stylized images with anime-style characteristics while preserving content structure. Validated outputs with side-by-side OpenCV comparisons and quality metrics.

CIFAR-10 Image Classification with Transfer Learning
Technologies: PyTorch, Transfer Learning, CNN
Timeline: 2024

Objective: Develop high-accuracy image classification model for CIFAR-10 dataset using transfer learning.

Approach: Developed and fine-tuned CNN model with dropout regularization and batch normalization for improved generalization. Applied transfer learning techniques to leverage pre-trained features. Conducted systematic hyperparameter tuning including learning rate, batch size, weight decay, and architectural parameters. Implemented comprehensive validation and testing protocols.

Results: Achieved 87% validation accuracy through careful optimization and architectural refinement.

Key Techniques: Batch normalization for training stability; dropout for regularization; transfer learning for reduced training time; systematic hyperparameter optimization.

Micromouse Maze Solver
Technologies: C++, Teensy 4.1, VL53L0X ToF Sensors, IMU, Flood-Fill Algorithm
Timeline: 2024 | IEEE BITS Pilani

Objective: Design and implement autonomous robot capable of solving complex mazes with real-time feedback.

Approach: Designed comprehensive embedded systems solution with Teensy 4.1 MCU, VL53L0X time-of-flight sensors, and 9-axis IMU. Implemented flood-fill maze-solving algorithm with PID control for precise motor control and lane-centering. Engineered sensor fusion for accurate orientation tracking and distance measurements. Developed shortest-path optimization post-maze-exploration.

Hardware Integration: Sensor fusion combining ToF range data and IMU orientation for robust position estimation; real-time motor control with PID tuning for precise movement.

Results: Successfully navigated complex maze structures autonomously with optimized shortest-path execution. Demonstrated robust hardware-software integration and real-time embedded systems design.

Template Adapted From: Siba Smarak Panigrahi