Selected work
Software, financial data and applied machine learning.
A selection of applications, experiments and work in progress.
01 / SOFTWARE ENGINEERING
A RAG-based customer support application with grounded policy answers, order-status lookup and explicit human escalation.
Deployed on ECS Fargate using Docker, an Application Load Balancer and Amazon RDS. GitHub Actions uses OIDC-based AWS access; CloudWatch, WAF and automated smoke tests support operations.
02 / FINANCIAL MACHINE LEARNING
A Python pilot study aligning FNSPID headlines with yfinance prices across six US technology equities, producing 40,755 labelled samples for next-day direction prediction.
Compared TF-IDF and Logistic Regression with direct FinBERT sentiment-label mapping using a chronological hold-out set. Analysed neutral-label bias and the limits of treating sentiment as a direct next-day price signal.
03 / FINANCIAL RESEARCH
An ongoing prototype combining financial fundamentals and dividend-specific sentiment to classify increases, no change and decreases.
NVIDIA (NVDA) is the initial case study. Mathematical models and a demonstration website are planned deliverables; the research framework is available as a draft.
04 / COMPUTER VISION
An empirical study of backbone and decoder choices in CellViT for nuclei instance segmentation and classification.
Compared ViT-256/SAM-H backbones and HoVer-Net/CPP-Net decoders. Reproduced PanNuke benchmarks and evaluated on 14 MoNuSeg images without fine-tuning; SAM-H showed the more consistent effect across datasets and metrics.
05 / DEEP LEARNING
PyTorch experiments spanning CNN classification, transfer learning and image-caption generation.
Applied augmentation, dropout and learning-rate tuning on TinyImageNet30; compared AlexNet fine-tuning strategies on CIFAR-10. Implemented an RNN caption decoder with ResNet features and evaluated pretrained CLIP/GPT-2 captioning models.
06 / BLOCKCHAIN
A Solidity project exploring peer-to-peer energy trading between community prosumers.
The project scope covers registration, Ether balances, recorder-controlled updates, supply–demand pricing and automated matching, with attention to validation, fairness and gas efficiency.