Production work, not portfolio fluff.

Retail and semi-professional traders lacked an all-in-one discipline system for position sizing, risk management, trade planning, and journaling. Existing tools focused on signal generation but ignored the execution and psychology side of trading.

Built a full-stack trading companion from the ground up — Vite + React + TypeScript frontend with a Supabase backend, Stripe subscription billing, and PWA support for mobile use. Automated tests and monitoring ensure production reliability.

Live SaaS product with Stripe-integrated payments, automated test suite, production monitoring, and a clean PWA experience. Four interconnected modules — position sizing, risk and discipline tools, plans and strategies, and trade history — form a complete trading workflow.

A regional healthcare system in a developing market lacked reliable case forecasting to plan ICU capacity, oxygen supply, and public-health interventions. Off-the-shelf global models did not generalise to local epidemiological data.

Led the machine-learning track in a globally distributed team of 15+ contributors. Implemented XGBoost time-series models with feature engineering on public-health data, and built an interactive dashboard for non-technical stakeholders to explore forecasts.

Model evaluation metrics (MAE, RMSE, R²) validated on held-out data. Dashboard deployed for ongoing use by regional health planners.

Clinicians in a Nordic public health system needed rapid access to rehabilitation guidelines, clinical protocols, and best practices scattered across hundreds of unstructured documents. Manual search was slow, inconsistent, and pulled staff away from patient care.

Designed and deployed a RAG knowledge assistant grounded in the organisation's internal clinical documentation corpus. Built a retrieval-augmented generation pipeline that indexes documents into a vector store and serves sourced answers through a conversational interface. Prioritised data sovereignty — all processing within the client's healthcare environment.

Clinicians can query rehabilitation protocols in natural language and receive sourced, grounded answers in seconds. The system surfaces the exact source for every claim, supporting clinical confidence and auditability.

State-of-the-art emotion recognition models were too large and power-hungry for on-device mobile deployment. Running inference in the cloud introduced latency, privacy concerns, and dependency on connectivity.

Designed a multi-phase pipeline: model selection and benchmarking on FER-2013, fine-tuning on RAF-DB for improved real-world accuracy, then dynamic INT8 quantisation to shrink the model footprint for mobile CPUs. Built the pipeline with PyTorch and OpenCV.

A production-ready pipeline from research dataset to mobile-optimised model. The INT8 quantised variant runs efficiently on-device while retaining competitive accuracy, enabling privacy-preserving, offline emotion recognition.