Portfolio

Filter by technology

25 of 25 entries

• 01

  Concord — Expert Matching Platform

  Production platform for The College Collective, a UK Further Education peer-support network, replacing a manual spreadsheet workflow. Matches each support request to candidate industry experts across 37 capability areas using a three-signal ranker — SBERT (all-MiniLM-L6-v2) embeddings, TF-IDF cosine similarity, and categorical overlap — served from a FastAPI microservice behind a React / Express / PostgreSQL Nx monorepo. Built and maintained as the sole engineer.

  PythonFastAPISBERTTF-IDFReactTypeScriptExpressJSPostgreSQLKyselyNx

  Read entry →
• 02

  Intelligent Matching System (MSc Dissertation)

  MSc Artificial Intelligence dissertation (University of Salford). Research and prototype automating the candidate-identification stage of the College Collective's expert matching. Combines categorical overlap over a 37-label taxonomy, TF-IDF cosine similarity, and Sentence-BERT embeddings into one ranked list, evaluated against 320 historical matching decisions. Achieved Precision@10 of 48.44% and MRR of 0.186, alongside a documented audit of five sequential evaluation-integrity bugs with before/after metrics for each.

  PythonSentence-BERTSBERTTF-IDFscikit-learnNLPPandas

  Read entry →
• 03

  Sentiment Classification — Nigerian Pidgin English

  NLP module project (MSc AI, University of Salford). Compared three TensorFlow embedding strategies — a trainable token embedding, NNLM-50, and NNLM-128 with normalisation — for 3-class sentiment on the NaijaSenti Pidgin English corpus (10,556 tweets), handling class imbalance with class weights and stratified splits. The trainable embedding learned from the target corpus won (weighted F1 0.72) over both pre-trained NNLMs, showing that domain-specific embeddings beat general-purpose ones for low-resource creole text.

  PythonTensorFlowNLPscikit-learnPandas

  Read entry →
• 04

  Topic Modelling — Global AI News Coverage

  NLP module project (MSc AI, University of Salford). Applied LDA and NMF to a self-collected corpus of global AI news, using grid search over coherence (LDA) and reconstruction error (NMF) to choose K. Both models converged on 10 consistent themes — cross-model validation — with NMF producing sharper, more lexically distinct topics and running far faster, while LDA's probabilistic outputs suit downstream routing. LDA coherence 0.5273; NMF reconstruction error 28.70.

  Pythonscikit-learnNLPTopic ModellingLDANMF

  Read entry →
• 05

  Diabetes Risk Screening with Decision Tree and Random Forest

  Machine Learning & Data Mining module project (MSc AI, University of Salford). Decision Tree and Random Forest for diabetes risk screening on BRFSS-2015 (253,680 records). Threshold optimisation (0.50 → 0.40) prioritises recall over precision — a clinical trade-off argued explicitly: the tuned Random Forest catches 88% of at-risk individuals at balanced accuracy 0.73, cutting the false negative rate by 48%.

  Pythonscikit-learnRandom ForestDecision TreeSupervised LearningPandas

  Read entry →
• 06

  Behavioural Segmentation of Eating Habits Using DBSCAN and OPTICS

  Machine Learning & Data Mining module project (MSc AI, University of Salford). DBSCAN and OPTICS density-based clustering to segment individuals by eating habits, with PCA beforehand. Compared on cluster stability, noise handling, and hyperparameter sensitivity — neither needs a pre-set cluster count, an advantage on noisy real-world data.

  Pythonscikit-learnDBSCANOPTICSClusteringUnsupervised LearningPCAPandas

  Read entry →
• 07

  Twitter Sentiment Classification — TF-IDF vs Naive Bayes

  Machine Learning & Data Mining module project (MSc AI, University of Salford). TF-IDF + Logistic Regression vs Multinomial Naive Bayes for Twitter sentiment, with standard NLP preprocessing. Logistic Regression wins at 73.45% accuracy and F1 ≈ 0.74; TF-IDF's down-weighting of common terms beats Naive Bayes's raw-frequency assumption on social-media text.

  Pythonscikit-learnNLPTF-IDFLogistic RegressionText ClassificationPandas

  Read entry →
• 08

  ClinicalTrials.gov Analytics with Spark SQL

  Big Data Tools and Techniques module project (MSc AI, University of Salford). SparkSQL exploratory analysis of the full ClinicalTrials.gov registry (572,935 rows) on Databricks — a temporary view queried entirely in SparkSQL, handling null-heavy and pipe-delimited real-world data across study-type, condition, trial-duration (35.29 months via DATEDIFF), and Alzheimer's-trend questions.

  PythonPySparkApache SparkSparkSQLDatabricksBig DataPandas

  Read entry →
• 09

  Steam Game Recommender with Spark MLlib ALS

  Big Data Tools and Techniques module project (MSc AI, University of Salford). Collaborative-filtering recommender on Steam-200k (200,000 interactions, 12,393 users, 5,155 games) using Spark MLlib ALS in explicit mode, with a log1p transform for right-skewed play hours and a 24-run MLflow grid search over rank, regParam, and maxIter. Best RMSE 1.4632 (rank 40, regParam 0.5, maxIter 10), down from a 1.519 baseline.

  PythonPySparkApache SparkMLflowALSCollaborative FilteringDatabricksBig DataPandas

  Read entry →
• 10

  Plastic Classification — MobileNetV2 vs InceptionV3

  Deep Learning module project (MSc AI, University of Salford). Fine-tuned MobileNetV2 and InceptionV3 on a small self-collected image dataset (163 train / 71 validation) to classify HDPE against PET. MobileNetV2 outperformed InceptionV3 on precision, recall, and F1 for both classes; both reached at least 80% training accuracy after fine-tuning. The report analyses why the task is hard — plastics differ mainly by texture and transparency — and proposes concrete data-collection changes. Companion project to the YOLO detection study.

  PythonTensorFlowKerasTransfer LearningMobileNetV2InceptionV3

  Read entry →
• 11

  Plastic Object Detection — YOLOv8 vs YOLOv9

  Deep Learning module project (MSc AI, University of Salford). Trained and compared YOLOv8 and YOLOv9 (50 epochs) on a self-annotated dataset for HDPE, PET, and PP bounding-box detection. YOLOv8 was the more conservative, higher-precision model (mAP@0.5 0.3364, precision 0.5964) versus YOLOv9's noisier output (mAP@0.5 0.2669); confusion-matrix analysis pinned PET/PP confusion and background misclassification as a shared data-quality issue. YOLOv8 recommended for production. Companion project to the transfer-learning classification study.

  PythonYOLOv8YOLOv9Object Detection

  Read entry →
• 12

  Field Trial Training Platform

  Modernization of a niche canine field-trial training platform: redesigned the UI, added mobile responsiveness, and built a sponsorship feature to open a new revenue stream. Migrated the codebase to an Nx monorepo on React 19 / Next.js, consolidating front- and back-end deployment and cutting hosting costs. Featured as a Toptal case study.

  ReactNextJSNode.jsTypeScriptNx

  Read entry →
• 13

  Bluelines

  Bluelines is a business management solution for SMEs. Bluelines abstracts away the boring paperwork carried out by business owners on a daily basis, allowing them to focus on growing their business.

  ReactTypeScriptTanStack QueryPythonDjangoGraphene

  Read entry →
• 14

  TeachShare

  The world’s largest platform for Personalized Learning

  SolidJSTailwindCSS

  Read entry →
• 15

  Autumn AI

  Autumn AI is an early-stage startup based in Canada working on the future of preventative mental healthcare

  ReactTypeScriptNextJS

  Read entry →
• 16

  Mobile Flashcards

  A flashcard app for iOS and Android

  ReactReact NativeTypeScript

  Read entry →
• 17

  react-drag-drop-browser

  Simple, customizable React drag and drop component. Also comes with a file browser. It is highly configurable in terms of styling. Do checkout the docs on the Github repo.

  ReactJavaScript

  Read entry →
• 18

  Olympic medals

  Analysing olympic gold medals from a wikipedia dataset

  PythonPandas

  Read entry →
• 19

  Energy and GDP

  Analysing energy and GDP data of various countries

  PythonPandas

  Read entry →
• 20

  Hypothesis testing

  Testing a hypothesis via data analysis

  PythonPandas

  Read entry →
• 21

  GDP and Inflation by year

  Variation of GDP and Inflation rate of Nigeria and UAE between 2005 and 2016

  PythonPandasMatplotlib

  Read entry →
• 22

  Temperature highs and lows

  Plotting temperature highs and lows for every calendar day of the year in Abu Dhabi, UAE for the ten-year period, 2005 and 2014

  PythonPandasMatplotlib

  Read entry →
• 23

  Visualizing voting pattern

  The "yes" vote patterns in California Counties on Proposition 37 in 2012

  ArcGIS Desktop

  Read entry →
• 24

  Moving Valmeyer

  Map showing the location of the new and old Valmeyer towns

  ArcGIS Desktop

  Read entry →
• 25

  Visualizing ozone concentrations

  Map of Hourly Average Ozone Concentrations around California in 2010-2011

  ArcGIS Desktop

  Read entry →