Cameron Jones

Projects

Things I've built, from coursework to side projects.

Fitter: Running Fitness Analytics with Machine Learning

Solo project · Oct 2026 – Present

A Java web app that tells runners whether they are actually getting fitter, using a heart-rate model trained on ~97,000 real runs and a personal layer that learns each runner's own response.

  • Built with Spring Boot and PostgreSQL, ingesting 1 GB Apple Health XML exports with a StAX streaming parser and async jobs, never loading whole files into memory.
  • Trained a gradient-boosted model in Java with the Smile ML library on 28M ten-second segments from ~97,000 real runs, predicting heart rate with 11% lower MAE than a standard linear baseline.
  • Found and fixed a Simpson's paradox that made heart rate look 31% less sensitive to pace than it really was.
  • Built a per-user ridge-regression layer that cut heart rate MAE by 39% (12.5 to 7.7 bpm), predicting each run only from that runner's earlier runs, across 82 runners held out from all model training.
  • Trained a second Smile model on 9,828 hard efforts to predict race times with a median error of 3.2% (5 km) and 3.4% (10 km), then used the Daniels–Gilbert formula to build a pace planner for any distance.
  • Generated coaching reports with Llama 3.2 self-hosted on Ollama, validating 100% of its numbers in Java before display and cutting its prompt by 59% (2,260 to 930 tokens) to run fast on a CPU-only server.
  • Kept all health data on my own server, running Spring Boot, PostgreSQL (Flyway migrations) and Ollama in Docker Compose with no data sent to outside services, backed by 227 JUnit 5 tests covering every module.
JavaSpring BootPostgreSQLSmile MLOllamaDocker Compose

Europe 2100: European Demographics Web App

Frontend lead, 2-person team · Apr 2026 – Present

A full-stack platform that turns 18,000+ Eurostat and World Bank data points into an interactive picture of 65 years of European population change, with projections to 2100.

  • Co-building a full-stack platform with React, Python, FastAPI and pandas.
  • Lead the 7-page React frontend, building an interactive map of Europe with D3 and TopoJSON where users scrub through 65 years of data and compare up to 8 countries, with every view saved in a shareable URL.
  • Made the site feel instant by caching the full dataset in a shared store built on React's useSyncExternalStore and lazy-loading pages with Vite code splitting, so switching indicators needs zero API calls.
  • Building a Python scenario model that projects population and ageing 75+ years ahead to 2100, measuring its error and uncertainty and letting users test “what if?” assumptions such as fertility at 1.8.
ReactD3TopoJSONVitePythonFastAPIpandas

Flight Data Visualiser

Team leader & developer, 4-person team · Feb 2026 – Apr 2026

A Java desktop app in Processing that loads 563,737 real US flights (10 airlines, 369 airports) into interactive charts of delays, diversions, busy routes and hourly traffic.

  • Led a 4-person university team through 5 weekly milestones with Git branches and pull requests, so all 4 of us could build in parallel.
  • Implemented all 4 chart types (histogram, bar, scatter and pie) from scratch with Processing's graphics API, organising flight data with Java HashMaps and writing the largest share of the team's code.
  • Designed a search screen with 11 combinable filters and a dual-handle time slider, filtering 560,000+ flights live with a match count shown before any chart is drawn.
  • Applied object-oriented design to split the app into 20 classes (screens, charts, widgets).
JavaProcessingData visualisationGit

Probability of Rain Tomorrow: East vs West Ireland

In progress

STU22004 Applied Probability I group project, 3-person team · Nov 2026 – Present

Does rain today make rain tomorrow more likely, and is that persistence stronger on the west coast than the east? A study of long-running daily Met Éireann records from Dublin Airport and Valentia Observatory.

  • Estimates P(rain tomorrow | rain today) against P(rain tomorrow | dry today) to test whether wet days are independent.
  • Tests the difference with a chi-square test and a permutation (shuffle) test over thousands of reorderings of the days.
  • Compares wet days per week to a Binomial model and dry-spell lengths to a Geometric model, checking whether spells are really memoryless.
  • My part: comparing every result between Dublin and Valentia, using Bayes' theorem to ask whether rain in Kerry today predicts rain in Dublin tomorrow, and bounding the estimates with Chebyshev's inequality.
ProbabilityStatisticsHypothesis testingOpen data

Personal website

This site. Built with Astro, with light and dark themes and no client-side framework.

AstroTypeScriptCSS