Hi, I'm Gunbir Walia

This project implements a facial landmark detection system for precise face alignment, a core task in computer vision with applications in identity verification, emotion recognition, and AR/VR technologies.

• Preprocessed facial images by standardising resolution, normalising illumination, and reducing image noise to deliver consistent, high-quality inputs for model training.
• Extracted Scale-Invariant Feature Transform (SIFT) descriptors around candidate landmark regions to capture robust local texture features resistant to scale, rotation, and lighting variations.
• Implemented a cascaded ridge regression framework, where successive regression stages refined landmark predictions iteratively to improve accuracy.
• Assessed model performance using normalised mean error (NME) across diverse facial poses and expressions, demonstrating reliable landmark localisation and strong accuracy, even under challenging conditions.
• Analysed system limitations, including susceptibility to severe occlusions and extreme profile views, and proposed enhancements through data augmentation, deep feature extraction, and CNN-based architectures for improved generalisation.

This project implements a binary text classification system to distinguish between spam and legitimate messages, addressing real-world challenges in automated email and message filtering.

• Conducted comprehensive text preprocessing including lowercasing, removal of punctuation/special characters, tokenisation, stopword removal, and stemming to normalise and clean the dataset.
• Engineered TF-IDF vectorisation with both unigrams and bigrams, capped at the top 5,000 most frequent terms to balance contextual richness with computational efficiency.
• Trained and evaluated Multinomial Naïve Bayes and Logistic Regression classifiers, analysing trade-offs between speed, interpretability, and predictive performance.
• Achieved 96% accuracy with Naïve Bayes and 98% accuracy with Logistic Regression on the validation set, with the latter demonstrating superior precision, recall, and F1-scores.
• Investigated and documented failure cases, such as text obfuscation and contextually ambiguous non-spam messages, identifying limitations of surface-level n-gram models and opportunities for enhancement through deeper contextual embeddings.

This project explores the hypothesis that simple agents evolved using microbial genetic algorithms can coherently exhibit a range of intricate behaviours within a simulated environment. However, these behaviours degrade in performance or consistency when transferred to physical robots due to the reality gap between the real world and a controlled simulation.

• Designed and developed a simulated 2D environment in Python to evolve Braitenberg vehicles capable of exhibiting complex, animal-like behaviours using microbial genetic algorithms.
• Engineered agents with minimal cognitive architecture to display a rich repertoire of behaviours, including aggressor, coward, lover, explorer, shy, one-eyed phototaxis, and sequential phototaxis - each evolved through tailored fitness functions aligned with specific behavioural objectives.
• Modelled each agent’s genome as weighted sensor-to-motor connections with independent biases, enabling direct evolutionary optimisation of+ sensorimotor couplings.
• Transferred evolved genotypes to a physical robot using a Raspberry Pi Pico and MicroPython, analysing performance degradation due to the simulation-to-reality gap.
• Addressed real-world embodiment challenges such as sensor noise, hardware asymmetries, and directional sensory constraints, by adapting training conditions and motor bias strategies.
• Explored and critically analysed foundational principles of embodied cognition, active perception, and sensorimotor intelligence, highlighting how emergent behaviour arises from agent–environment interactions rather than internal complexity.

This project implements a comprehensive relational database that models a music ecosystem, featuring a detailed schema for artists, albums, songs, record labels and user-generated content like playlists, akin to the kind of databases you would find in the backend of music streaming platforms like Spotify or Apple Music.

• Created and rendered 3D assets in Blender and used Figma to design various UI elements.
• Managed project documentation including requirements analysis, risk management, testing plans and sprint documentation, ensuring clear traceability from client expectations to delivered features.
• Utilised Git & GitHub for version control and OneDrive for asset and documentation sharing, maintaining a smooth workflow and minimising redundancies.
• Coordinated sprint planning and task allocation using Gantt and PERT charts, following Agile methodology to deliver functional prototypes at the end of each sprint. Maintained detailed changelogs for stakeholder review, mirroring real-world development practices and fostering adaptability, problem-solving, and collaborative skills.

This project implements a comprehensive relational database that models a music ecosystem, featuring a detailed schema for artists, albums, songs, record labels and user-generated content like playlists, akin to the kind of databases you would find in the backend of music streaming platforms like Spotify or Apple Music.

• Engineered database performance by strategically adding indexes and primary/foreign key constraints to optimize query speed and ensure data integrity.
• Identified and documented key limitations of the design, addressing crucial scalability challenges (e.g., database sharding) and opportunities for future expansion.
• Developed a comprehensive suite of SQL queries to demonstrate core functionality and data manipulation.

This project develops an interpreter for a simple, custom-designed programming language using ANTLR (Another Tool for Language Recognition). It features a grammar specification with support for typed variable declarations, expressions, functions, and control flow constructs.

• Defined the language's syntax using context-free grammar, enabling ANTLR to automatically generate a parser for syntactic analysis.
• Constructed parse trees from user code and performed semantic analysis to ensure correctness, catching type mismatches, undeclared variables, and other logical errors.
• Implemented execution for valid code, alongside detailed error reporting to guide users in correcting invalid input.