AIMLSoftwareBackendEngineer

K-Means Market Basket Analysis

Two-person honours research project on speeding up association rule mining with a K-Means pre-filter that clusters items on revenue and co-purchased revenue and drops the lowest-value cluster before Apriori runs. Built Apriori from scratch in Java (candidate generation, pruning, support counting) as the supporting algorithm that mines rules from the filtered data. Benchmarked the K-Means-filtered pipeline against Apriori on the full dataset and a published clustering method on 130K+ records, with comparable top-100 rule value.

Transfer Learning & Domain Generalization

Group project (three people) on how layer freezing affects transfer learning. Ran 90 controlled fine-tuning experiments: ResNet-50 and EfficientNet-B3 across CIFAR-10, CIFAR-100, and brain tumor MRI, 3 data sizes, and 5 freeze depths. Built the PyTorch pipeline and layer-freezing logic, and reached 95.9% on brain tumor MRI.

Stock Market Analysis and Prediction

Compared LSTM and SVM models for S&P 500 closing-price regression and up/down classification on the CNNPred dataset. Engineered lag, return, moving-average, and volatility features, tuned the LSTM with grid search and the SVM across linear, polynomial, and RBF kernels, and evaluated with RMSE, MAE, R², accuracy, precision, recall, and confusion matrices.

Note: the final report lists 99.1% SVM accuracy. That run used same-day features, so it isn’t a true forecast. With that leak removed, next-day accuracy is about 56%.

MindForge AI Suite (StudyAI + ResearchAI)

Architected a fully local LLM platform using LLaMA 3.2 via Ollama, analyzing PDF, DOCX, PPT/PPTX, TXT, and Markdown documents with zero API cost. Engineered a RAG pipeline with TF-IDF retrieval over overlapping 400-word chunks, sending only the most relevant passages to the model to cut prompt context across multi-document research sessions. Parallelized summarization across 4 workers (ThreadPoolExecutor) with real-time SSE streaming, plus integrated web retrieval and automated APA citation generation.