ML & AI Projects

Production platform for The College Collective, replacing a manual spreadsheet process for matching support requests from Further Education colleges to industry experts across 37 capability areas in 7 categories. The ML layer is a FastAPI microservice serving SBERT (all-MiniLM-L6-v2) embeddings combined with TF-IDF cosine similarity and categorical overlap scoring to produce a ranked shortlist of candidate experts per support request. The application layer is a React / TypeScript / Express / PostgreSQL Nx monorepo on Render, with Resend transactional email, a pg-boss job queue, and Sentry + LogRocket observability.

Python · FastAPI · SBERT (all-MiniLM-L6-v2) · TF-IDF · React · TypeScript · Express · PostgreSQL · Kysely · Nx monorepo · pg-boss · Resend · Sentry · LogRocket

Research and prototype automating the candidate-identification stage of the College Collective's expert matching. Three complementary signals — categorical overlap over a 37-label taxonomy, TF-IDF cosine similarity over free-text descriptions, and semantic similarity via Sentence-BERT — are combined into a single ranked list, evaluated against 320 historical matching decisions from four years of programme data. Substantial effort went into an honest evaluation baseline: five sequential data-quality and pipeline-integrity bugs were identified and documented with before/after metric impact, including a request-identity collision, an empty-text SBERT artefact, and inconsistent ground truth.

Python · Sentence-BERT · TF-IDF · scikit-learn · pandas · NumPy · FastAPI · React

Comparative study of three TensorFlow embedding strategies for sentiment classification of a low-resource creole corpus: a trainable token-based embedding, NNLM-50, and NNLM-128 with text normalisation, on the NaijaSenti Pidgin English dataset (10,556 tweets; positive / neutral / negative). Class imbalance was handled with class weights and stratified re-splitting. The central finding is that a trainable embedding learned from the target corpus outperforms both pre-trained NNLM models — a domain mismatch between Google News-trained NNLMs and informal creole-register Twitter text. NNLM-128's punctuation normalisation reduces the OOV rate and partially mitigates this.

Python · TensorFlow · TensorFlow Hub · HuggingFace Datasets · scikit-learn · pandas

Applied LDA and NMF to a self-collected corpus of global news covering artificial intelligence, identifying latent thematic structure and comparing both models qualitatively and quantitatively. Grid search over coherence (LDA) and reconstruction error (NMF) selected K. NMF produces sharper, more lexically distinct topics (e.g. surfacing “anthropic, openai, chatgpt, claude” in a company-research topic); LDA shows more word-level mixing but richer probabilistic outputs suited to downstream tasks such as document routing. Per-topic interactive pyLDAvis visualisations were generated.

Python · scikit-learn · Gensim · pyLDAvis · NLTK · pandas · matplotlib

Applied Decision Tree and Random Forest classifiers to the BRFSS-2015 dataset (253,680 records) for diabetes risk screening. The core challenge was clinical: minimising false negatives matters more than raw accuracy when screening for disease. Applied threshold optimisation — shifting the classification boundary from the default 0.50 to 0.40 — to raise recall at the cost of precision, a deliberately argued clinical trade-off. Both models reached balanced accuracy of 0.73; the threshold-tuned Random Forest caught 88% of at-risk individuals (up from ~65% at the default threshold), cutting the false negative rate by 48%.

Python · scikit-learn · pandas · NumPy · matplotlib · Jupyter

Applied DBSCAN and OPTICS density-based clustering to segment individuals by eating habits, using PCA for dimensionality reduction prior to clustering. Both algorithms were compared on their ability to identify meaningful behavioural segments without a pre-specified number of clusters — an advantage over centroid-based methods for noisy real-world datasets. The analysis evaluated cluster stability, noise-point sensitivity, and the effect of DBSCAN's epsilon and min_samples hyperparameters, contrasted with OPTICS's reachability-based approach.

Python · scikit-learn · pandas · NumPy · matplotlib · Jupyter

Compared TF-IDF + Logistic Regression against Multinomial Naive Bayes for sentiment classification of Twitter text. The pipeline covered standard NLP preprocessing — tokenisation, stopword removal, vectorisation — before model training. TF-IDF with Logistic Regression outperformed Naive Bayes on accuracy and F1, the argued reason being that TF-IDF's ability to down-weight frequent but uninformative terms beats Naive Bayes's raw term-frequency assumption on social-media text.

Python · scikit-learn · NLTK · pandas · matplotlib · Jupyter

SparkSQL exploratory analysis of the full ClinicalTrials.gov registry (572,935 rows, 14 columns) on Apache Spark / Databricks — registered as a temporary view and queried entirely in SparkSQL. Handled real-world data quality issues: 68% null Collaborators, 71% null Acronyms, and pipe-delimited multi-value Conditions requiring SPLIT and EXPLODE. Four questions: study-type distribution, top-10 conditions by frequency, mean trial duration via DATEDIFF, and the Alzheimer's completed-trial trend 1996–2026, converted to pandas and plotted with matplotlib.

Python · PySpark · Apache Spark · SparkSQL · Databricks · pandas · matplotlib

Collaborative-filtering recommendation engine on the Steam-200k dataset (200,000 interactions, 12,393 users, 5,155 games), built on Apache Spark / Databricks. EDA showed heavily right-skewed play hours (mean 48.88h, median 4.5h, max 11,754h), motivating a log1p transform before training. Trained Spark MLlib ALS (Alternating Least Squares) in explicit mode with integer-indexed game IDs, tuned by a 24-run MLflow grid search over rank ∈ [5, 10, 20, 40], regParam ∈ [0.01, 0.1, 0.5], and maxIter ∈ [5, 10]. Produces top-5 recommendations per user, with game names decoded from the StringIndexer mapping.

Python · PySpark · Apache Spark · MLlib (ALS) · MLflow · StringIndexer · Databricks · pandas · NumPy

Comparative study of two transfer-learning architectures for plastic-type classification (HDPE and PET) on a self-collected image dataset. Images were gathered, labelled, and fine-tuned on both MobileNetV2 and InceptionV3 with TensorFlow/Keras. The small dataset (163 training, 71 validation) was a genuine constraint — both models learned but with inconsistent training curves. MobileNetV2 outperformed InceptionV3 on precision, recall, and F1 for both classes. The report analyses why the task is hard — plastics differ mainly by texture and transparency — and proposes concrete data-collection changes.

Python · TensorFlow · Keras · MobileNetV2 · InceptionV3 · matplotlib · Google Colab

Trained and compared YOLOv8 and YOLOv9 on a self-annotated dataset for detecting HDPE, PET, and PP with bounding boxes, over 50 epochs. Both showed consistent training-loss reduction but volatile validation metrics, consistent with limited data. YOLOv8 behaves conservatively — higher precision, fewer but cleaner detections; YOLOv9 is more aggressive with noisier predictions. Confusion-matrix analysis identified PET and PP as the most confused classes, with background misclassification the dominant failure mode for both — a shared data-quality issue rather than a model-specific one.

Python · Ultralytics YOLOv8 · YOLOv9 · OpenCV · Google Colab · Roboflow