Master Project and Skills Inventory

• Engineered an end-to-end non-destructive plant-phenotyping pipeline to estimate lettuce dry-shoot biomass from RGB-D imagery.
• Restored raw sensor depth maps via Region-of-Interest (ROI) cropping, OpenCV inpainting, bilateral filtering, and depth-range thresholding.
• Implemented DeepLabV3+ with a ResNet34 backbone and Atrous Spatial Pyramid Pooling (ASPP) for high-precision plant canopy segmentation.
• Optimized network using Tversky loss (weighting false negatives), spatial data augmentations, early stopping, and IoU/F1 metric tracking.
• Extracted 39 handcrafted RGB, depth, geometric, 3D voxel volume, convex hull, vegetation indices (ExG, NDVI equivalents), and morphological shape features.
• Trained and benchmarked regression models: LightGBM, Gradient Boosting, Huber regression, staged prediction, cross-validation, and ensemble voting.
• Built cross-platform workflows supporting RGB, depth, masks, CSV metadata, Apple Silicon MPS acceleration, and CPU fallback.

• Curated and processed Mozilla Common Voice English data spanning 13 distinct global accent groups.
• Built an audio engineering pipeline: SNR filtering, duration filtering, long recording segmentation, 16 kHz mono resampling, and amplitude normalization.
• Applied acoustic data augmentations including Gaussian noise injection, pitch shifting, and silence insertion.
• Analyzed prosody and phonetic features: MFCCs, fundamental frequency (F0/pitch), RMS energy, spectral centroids, PCA, and t-SNE clustering.
• Fine-tuned Facebook's wav2vec2-base via self-supervised masked frame prediction and contrastive InfoNCE loss.
• Trained a 768-to-512-dimensional bottleneck accent adapter network using Mean Squared Error (MSE) loss.
• Integrated SpeechT5 and SpeechT5 HiFi-GAN neural vocoder for accent-conditioned speech synthesis.
• Engineered robust Kaggle NVIDIA P100 training scripts featuring checkpoint rotation, resume capabilities, deterministic seeds, and device detection.
• Synthesized speech across all 13 accent profiles and conducted a qualitative listening evaluation.

• Architected a full 3-stage Reinforcement Learning from Human Feedback (RLHF) pipeline: Supervised Fine-Tuning (SFT), Reward Modeling (RM), and Policy Optimization.
• Conducted a head-to-head empirical comparison of Proximal Policy Optimization (PPO) versus Direct Preference Optimization (DPO) on GPT-2 and GPT-2-Medium.
• Utilized Stanford SST-2 sentiment and UltraFeedback pairwise preference datasets.
• Engineered reward models featuring linear scalar heads and partially frozen transformer layers, trained via binary cross-entropy with warmup and cosine learning-rate schedules.
• Implemented PPO with Generalized Advantage Estimation (GAE), clipped surrogate policy objectives, dynamic KL-divergence penalties, and value heads.
• Implemented DPO directly optimizing policy parameters from preference pairs without training a separate reward model.
• Conducted local training on Apple Silicon Metal Performance Shaders (MPS).
• Evaluated reward model accuracy, pairwise preference accuracy, language model perplexity degradation, reward hacking phenomena, and blind peer preference.

• Designed a reproducible restoration-to-detection benchmark comparing classical Richardson–Lucy deconvolution against learned NAFNet (Nonlinear Activation Free Network).
• Trained and validated on GoPro sharp/blurred image pairs and synthetic motion blur models generated from COCO-2017.
• Fine-tuned YOLOv8n specifically for degraded and restored imagery to quantify the downstream benefits of image restoration.
• Benchmarked PSNR, SSIM, detection confidence distributions, bounding-box counts, blur severity gradients, mAP@0.5, and failure modes.
• Utilized Apple Silicon MPS/CPU evaluation alongside cloud Kaggle Tesla T4 GPU training with fixed seeds and strict splits.

• Engineered a full-stack decentralized NFT marketplace web application connecting smart contracts with a modern frontend.
• Built backend and smart contracts using Solidity, Hardhat, Ethereum Sepolia testnet, MetaMask wallet integration, Node.js, and IPFS/database storage.
• Implemented core Web3 workflows: ERC-721 token minting, ownership verification, marketplace listings, escrow transactions, and balance management.
• Developed an AI-driven search engine featuring Natural Language Processing (NLP) with part-of-speech (POS) tokenized search.
• Researched and specified computer vision image retrieval using visual feature embeddings to detect duplicates and enable reverse-image NFT search.
• Designed dark-themed UI/UX featuring product zoom, CSS micro-animations, gamification elements, and educational Web3 onboarding.
• Authored a comprehensive conference paper and a formal patent-style draft specification for multimodal digital asset retrieval.

• Designed and fabricated a two-wheeled self-balancing social robot combining inverted pendulum control with computer vision and voice interaction.
• Hardware architecture: Raspberry Pi Zero W (companion computing), ESP32-CAM (video streaming), MPU6050 6-DOF IMU, high-torque DC gear motors, L298N driver, Li-ion battery regulation, servo neck, IR sensors, microphone, speaker, and LCD display.
• Developed closed-loop Proportional-Integral-Derivative (PID) control firmware in Arduino C++ to maintain vertical stabilization under dynamic perturbations.
• Implemented Raspberry Pi Python service executing OpenCV face and object tracking, real-time video streaming, and voice command processing.
• Integrated Google Speech-to-Text, Google Search API, gTTS, and eSpeak to deliver conversational answers, weather forecasts, and home automation triggers.
• Built mobile and web dashboard prototypes to stream live camera telemetry and control manual robot navigation.
• Co-authored and published technical paper at the ICSTCEE conference.

• Investigating mathematical optimization of sensor modality, quantity, and spatial placement across municipal wastewater networks to proactively detect chokes, overflows, and blockages.
• Conducted an extensive systematic literature review synthesizing 50+ peer-reviewed papers into an interactive, searchable digital library with metadata filters and summaries.
• Formulated a comparative trade-off matrix covering 16+ sensor types: radar, ultrasonic, LiDAR, ToF, float, tilt, pulse, conductivity, temperature, pH, turbidity, UV/Vis, fluorescence, H₂S gas, acoustic, and CCTV.
• Analyzed real-world operational constraints: wireless telemetry (LoRaWAN/cellular), battery lifecycle, biofilm fouling, sensor calibration drift, physical manhole access, hazardous maintenance, redundancy, and capital vs operational expenditure (CapEx/OpEx).
• Formulated a mathematical sensor-placement optimization model on network graphs using binary decision variables, information entropy, mutual information, Fisher information matrices, source distinguishability, and detection latency.
• Investigated combinatorial solvers: greedy submodular optimization, NSGA-II/III multi-objective genetic algorithms, network centrality heuristics, and Mixed-Integer Linear Programming (MILP).
• Produced client presentations, project charters, Paper 1, and Paper 2.

• Architected an interactive music streaming web application built with React, Vite, and modern Web Audio APIs.
• Engineered a custom studio-deck vinyl turntable interface featuring realistic vinyl scratching, tonearm tracking physics, drag-to-seek, rotational inertia, mechanical brake decay, RPM tempo adjustments, and custom keyboard sliders.
• Integrated Dexie.js / IndexedDB client-side database, lazy loading, service workers for offline caching, and native Node test suites.
• Built a headless media ingestion pipeline on macOS utilizing yt-dlp, FFmpeg audio transcoding, browser cookies, and metadata tagging.
• Engineered robust title-matching regex algorithms to preserve authentic video titles while filtering out low-quality re-uploads and unrelated music videos.
• Implemented dynamic artwork extraction, taste profile generation, audio embedding vectors, instant search, queue management, and automated browser smoke tests.

• Log4Shell Exploit Analysis (CVE-2021-44228): Authored an exhaustive technical dissection of the Apache Log4j JNDI/LDAP remote code execution (RCE) vulnerability. Documented exploit chains, affected assets, root causes, attacker attribution, detection signatures, WAF regex rules, outbound LDAP blocking, SOC playbooks, SBOMs, and defence-in-depth principles.
• AgeWell Innovations Risk Assessment: Conducted an enterprise security risk assessment for an AI-enabled aged-care digital transformation involving remote telemetry and unmanaged BYOD endpoints.
• Constructed a formal 5x5 Likelihood vs. Impact risk matrix identifying 5 critical attack scenarios.
• Designed mitigation architecture: Mobile Device Management (MDM), application containerization, Role-Based Access Control (RBAC), Multi-Factor Authentication (MFA), Privileged Access Management (PAM), AES-256 encryption at rest/transit, zero-trust VPNs, binary allowlisting, phishing simulation, and phased compliance governance.
• Mapped controls to Australian Privacy Principles (APPs), Notifiable Data Breaches (NDB) scheme, ACSC Essential Eight, NIST Cybersecurity Framework, and ISO/IEC 27001.

• Co-designed a proactive automated surveillance system concept for detecting abandoned luggage and unattended objects in transport terminals and airports.
• Formulated a vision pipeline combining adaptive background reference-frame differencing with Region-of-Interest (ROI) segmentation to isolate static items.
• Designed integration with deep object detectors (YOLO, Faster R-CNN, SSD) paired with multi-object centroid tracking to compute unattended idle duration.
• Defined tiered automated alert triggers based on stationary dwell time thresholds and security operator escalation pathways.
• Produced system architecture specifications, requirements analysis, academic literature reviews, and engineering design posters.

• Formulated early product strategy, market research, and technical architecture for a B2B Manufacturing-as-a-Service (MaaS) and managed supply-chain marketplace.
• Conducted market research across four key verticals: home décor, commercial construction, healthcare supplies, and educational institutions.
• Architected platform modules: automated supplier discovery, demand aggregation algorithms, quality-assurance inspection workflows, multi-modal logistics routing, CAD design replication verification, managed procurement milestones, and trade finance/escrow.
• Authored investment pitch decks, financial unit economics models, and customer validation interviews as founding technical lead.

• Designed a 3-page portrait digital and print culinary zine workflow focused on preserving regional Telugu and multilingual family culinary heritage.
• Enforced strict cultural and culinary safeguards to preserve exact verbatim voice transcripts, prevents hallucinated ingredients, and protects authentic measurements.
• Architected workflow pipeline: raw voice transcription, bilingual Telugu-to-English translation, culinary photo styling, hand-drawn step illustrations, process infographics, nutritional estimates, and cultural origin notes.
• Built a responsive hosted web application using Next.js / Vinext featuring page-turn navigation, editorial typography, sample loading, and print-to-PDF export styling.

Curated a highly constrained small-scale image dataset. Filtered low-information, corrupted, and unreadable images using OpenCV Canny edge-density filtering. Retained 3 clean butterfly classes. Trained a compact PyTorch CNN with batch normalization, dropout, and spatial augmentations on Apple MPS. Benchmarked Adam vs SGD optimizers, achieving ~96% validation accuracy and 30/30 (100%) accuracy on a held-out test split.

Trained an MLP on the Pima Indians Diabetes dataset. Applied strict train-only median imputation and standardization to prevent data leakage. Engineered age bins, BMI-age interaction terms, and log-transformed insulin features. Benchmarked Small, Medium, and Wide architectures with BatchNorm, ReLU, Dropout, Kaiming initialization, and BCEWithLogitsLoss with positive class weighting to prioritize clinical recall over raw accuracy.

Built models to classify 8 distinct categories of blood cells. Audited the dataset to eliminate corrupted images, zero-variance frames, and perceptual-hash (pHash) near-duplicates. Applied CLAHE preprocessing. Constructed SimpleCNN, ImprovedCNN, and custom ResNet architectures from scratch. Benchmarked against HOG feature baselines with SVM and Logistic Regression. Tuned AdamW, cosine annealing schedules, label smoothing, and gradient clipping.

Constructed a convolutional VAE in PyTorch. Implemented Gaussian reparameterization ($\mu + \sigma \odot \epsilon$) over a 16-dimensional latent space. Formulated loss using Binary Cross-Entropy reconstruction and KL divergence with monotonic/cyclical annealing. Utilized ConvTranspose2d, SiLU activations, and BatchNorm layers. Trained 20 epochs on Apple Silicon MPS. Visualized 2D spherical latent interpolations and generated crisp synthetic digit grids.

Trained PyTorch MLPs and CNNs on 48x48 FER2013 facial data. Held out PublicTest and PrivateTest splits. Used data augmentation, cosine scheduling, and macro-F1 early stopping. Highlighted fundamental failure modes on minority classes: MLP reached ~0.37 accuracy, while CNN reached ~0.55 accuracy with macro-F1 around 0.43. Authored an ethical impact assessment documenting why facial emotion recognition should not be deployed in high-stakes human decision-making.

Benchmarked OpenAI CLIP ViT-B/16 and ViT-B/32 zero-shot classification on MNIST with prompt engineering and t-SNE embedding visualizations. Trained linear and residual MLP cross-modal adapters over frozen CLIP image/text embeddings for Flickr30k retrieval (evaluating bidirectional Recall@K). Built a Deep Convolutional GAN (DCGAN) on Fashion-MNIST with transposed convolutions and LeakyReLU.

Executed comprehensive data restructuring using R, tidyverse, and Quarto. Ingested messy international cricket batting records and converted from wide to long format. Extracted batting orders, runs, balls faced, teams, and player roles. Cleaned and structured board-game datasets. Constructed publication-quality ggplot2 visualizations including faceted histograms, boxplots, and scatter distributions. Applied non-parametric Wilcoxon rank-sum hypothesis testing to compare game performance distributions.

Analyzed market revenue dynamics across ~7,700 commercial board games. Compared raw and log-transformed revenue distributions. Built and evaluated ordinary least squares (OLS) linear and log-linear econometric models. Evaluated regression diagnostics (residual normality, heteroskedasticity, leverage points), computed prediction intervals, and interpreted economic rating coefficients for commercial strategy.

Constructed a full predictive pipeline in R using tidymodels. One-hot encoded game categories, defined high-rating binary targets, eliminated data-leakage variables, and enforced a stratified 75/25 train/test split. Formulated recipe workflows and hyperparameter grids for K-Nearest Neighbors (KNN) and Decision Tree classifiers with 5-fold cross-validation tracking accuracy and ROC-AUC.

• SIEM Operations & Threat Hunting: Forensic investigations using Sysmon and Splunk. Analyzed rundll32.exe execution paths, malicious Scheduled Tasks, correlated Sysmon Event ID 129 with Windows Security Event 4688, and detected obfuscated PowerShell commands.
• Host Intrusion Detection (HIDS): Deployed Wazuh server/agent infrastructure and OSSEC manager. Configured File Integrity Monitoring (FIM), custom rule IDs, and monitored SSH brute-force bursts.
• Network Reconnaissance: Authorized SNMP community string enumeration using Nmap and snmp-brute scripts in isolated virtual laboratories.
• Threat Intelligence & MITRE ATT&CK: Mapped real-world cyber incidents to MITRE ATT&CK techniques, including T1218.011 (Signed Binary Proxy Execution: Rundll32), Log4Shell (CVE-2021-44228), MOVEit Transfer SQLi (CVE-2023-34362), Storm-0978 RomCom malware, Heartbleed, Shellshock, Pass-the-Hash (PtH), Process Hollowing, and Utilman.exe privilege escalation backdoors.
• Security Frameworks: Designed risk registers, Likelihood/Impact matrices, and mitigation plans incorporating MDM, RBAC, MFA, PAM, zero-trust VPNs, AES-256 encryption, WAF rules, and SBOMs mapped to NIST CSF, Essential Eight, and ISO 27001.

Prototyped a two-wheeled balancing platform using Arduino Uno/Nano, MPU6050 accelerometer/gyroscope, HC-05 Bluetooth module, and DC gear motors. Developed closed-loop PID balancing firmware in the Arduino IDE to stabilize inverted pendulum dynamics while accepting directional commands via Bluetooth.

• ROS & TurtleBot3: Programmed ROS / rospy nodes to publish velocity commands and navigate TurtleBot3 mobile bases.
• FastAPI Robot Services: Built high-performance FastAPI endpoints receiving QR code JSON payloads to trigger immediate physical robot actions and audio announcements.
• Edge Compute & Vision: Integrated NVIDIA Jetson Xavier NX and Raspberry Pi 4 with Intel RealSense D435 RGB-D depth cameras, implementing depth-to-color alignment and real-time ONNX runtime object detection.
• Actuation & Control: Interfaced OpenCR microcontrollers over serial and controlled Dynamixel smart servos via the Dynamixel SDK, executing multi-threaded visual tracking and motor angle calculations.
• IoT Microcontrollers: Deployed ESP32-CAM video streaming servers over HTTP, interfaced Raspberry Pi GPIOs, and evaluated TensorFlow Lite inference latency on embedded ARM chips.

• Worked as UI/UX Intern (Sept–Nov 2023) collaborating with cross-functional design and development teams.
• Conducted user research, stakeholder interviews, user journey mapping, and persona synthesis.
• Created wireframes, high-fidelity mockups, and interactive clickable prototypes using Figma.
• Executed usability testing, cognitive walkthroughs, and iterated on user feedback loops.
• Applied Design Thinking, WCAG accessibility standards, information architecture, microinteractions, and design systems.

Developed an interactive React application for English learning with Web Speech API voice-controlled quiz navigation, restart logic, and Ngrok tunnels. Engineered advanced modules: WebRTC microphone capture, Axios REST integration, touch drag-and-drop, Recharts dashboards, Intro.js product tours, Fuse.js fuzzy search, and media controls.

Conducted an in-depth PMBOK failure analysis of Apple's cancelled AirPower mat. Evaluated scope creep, thermal engineering limits, stage-gate breakdowns, risk registers, value engineering, go/no-go thresholds, and sustainability impacts.

Served as Project Manager for Group 73. Authored Project Charter, Work Breakdown Structure (WBS), milestone schedules, Stakeholder Register, Power-Interest Grid, comprehensive Risk Register with qualitative scoring, communications matrix, escalation pathways, and document governance.

• AI vs. Performers & Likeness Rights: Poster on synthetic media in entertainment: voice/likeness rights, digital scanning, performer displacement, posthumous consent, SAG-AFTRA protections. Evidence: Poster Cons3.pdf
• AI Creativity & Meme Ethics Case Study: Evaluated generative AI creativity via Deontological, Utilitarian, and Virtue Ethics. Analyzed copyright infringement in training corpora, artist consent, compensation models, C2PA cryptographic watermarking, and AI labeling. Evidence: Cadmus Final Submission
• Whistleblower Governance Report: Addressed unlabelled AI content, lack of training consent, unfair competition, independent auditing committees, and phased compliance governance.
• Algorithmic Justice in Public Welfare: Analyzed automated welfare systems (Robodebt context): human-in-the-loop oversight, disparate impact, algorithmic impact assessments, appeal rights, distributive justice.
• Cognitive & Implicit Bias Mitigation: Structured practices separating empirical observation from subconscious assumption, incorporating counter-stereotyping and empathetic communication.

Contributed directly to corporate UI/UX client deliverables. Collaborated with senior product designers and front-end engineering teams. Formulated user research protocols, journey maps, and high-fidelity Figma prototypes. Conducted usability testing and implemented WCAG 2.1 accessibility standards across web platforms.

Contributed to applied Data Science and AI/ML initiatives while spearheading technical demonstrations and product pitches for commercial robotics systems. Delivered live technical presentations to prospective clients and synthesized audience feedback into actionable engineering requirements. (Honest boundary: Frame as technical demonstration and AI/ML contributor; do not claim proprietary ownership of Vistan commercial robots).

Completed intensive web application development training and project work covering HTML5, CSS3, JavaScript, AngularJS, Bootstrap, PHP, MySQL, jQuery, WAMP/XAMPP environments, full SDLC, and browser compatibility testing. (Note: Verify completion certificates match personal records before formal resume inclusion).

Delivered front-line customer engagement, technical advisory, and commercial sales across premium consumer electronics (smartphones, computing hardware, audio engineering systems, and smart-home IoT devices). Explained warranty terms, hardware specifications, and post-sales technical support.

• AUSA / YouX Crew Volunteer: University of Adelaide student community volunteer. Managed logistics, uniform-replacement pop-ups, and student campus engagement programs.
• FORCE Club Event Photographer: Lead photographer documenting large-scale university events, workshops, and student festivals in dynamic, fast-paced environments.
• Online Creative Portfolio: Published professional UI/UX and product design cases on Behance Portfolio Profile.