Portfolio

I’m an analytics engineer and full-stack AI product builder. Professionally I build production data platforms - dbt transformation layers, Airflow / Google Cloud Composer orchestration, BigQuery warehousing and custom REST-API integrations across a couple of dozen advertising-platform sources. Alongside that I design and ship full-stack, AI-powered products end to end. I came up through data science and machine learning, so those projects are here too.

Products & Engineering

Analytics Engineering (dbt · Airflow · BigQuery)
DnDwithToph.com
DnDwithToph Exploratory Data Analysis
Zombie Survival Game

AI / Data Science

Fine-tuning LLM for RPG Statblocks
Handwritten Character Recognition / Automated Scorecard Processing
Smoker vs. Non-Smoker Classification
Red Card / Goal Exploratory Data Analysis
EIFO Data Extraction
Electric Vehicle kWh Consumption Forecasting
Bitcoin Price Forecasting
Structured Medical Text Classification with NLP
Transfer Learning in Food Image Classification
Dungeons & Dragons Race Classification using Scikit-learn

Featured product

History Check - Your campaign never forgets.

A web-first app for tabletop-RPG Game Masters that records a session, transcribes it, and uses AI to do the bookkeeping - generating summaries and automatically extracting the NPCs, locations, items, monsters and organisations that came up, then tracking them across an entire campaign. Built solo, end to end.

Frontend: Flutter web (Dart), Riverpod state management, go_router, responsive layout.
Audio capture: in-browser recording via the MediaRecorder API (Opus), chunked and uploaded incrementally so a GM can close the browser mid-session and lose nothing; crash/refresh recovery reconciles the local cache against cloud storage.
Backend: Supabase / PostgreSQL as the single source of truth via PostgREST, with row-level security enforcing a multi-user owner/viewer (GM/player) permission model, token-based invite links, and SECURITY DEFINER RPCs.
AI pipeline: a fully server-side pipeline on Deno edge functions - audio → AssemblyAI transcription (speaker diarization) → Google Gemini for structured (JSON-schema) summarisation and entity extraction. API keys stay server-side.
Production engineering: Sentry observability, Supabase Realtime status updates, a Paddle/Stripe billing & feature-gating layer, separate staging/production environments and deploy pipelines, and 1,100+ Flutter tests plus ~70 edge-function test files.

Flutter · Dart · Supabase · PostgreSQL · Deno · TypeScript · AssemblyAI · Google Gemini · Next.js · Sentry · Paddle

History Check app: the AI Historian answering a campaign question alongside auto-extracted NPCs, locations and items

The in-app AI “Historian” answering a question about the campaign, beside the world database of auto-extracted entities.

Products & Engineering

Analytics Engineering - dbt · Airflow · BigQuery

Professional work building and maintaining production data platforms on Google Cloud.

Orchestrated data pipelines on Airflow / Google Cloud Composer, running both dbt transformations and custom business-logic DAGs.
Ingested from ~two dozen advertising-platform sources through custom REST-API integrations, landing raw data into BigQuery.
Built and tested modelled, business-ready data with dbt (staging → marts), with schema/data tests guarding correctness.
Worked across the stack in Python 3.11 and SQL, plus agentic-AI tooling layered on top of the data platform.

DnDwithToph.com

Flask framework with SQLAlchemy database management
User Authentication and Session Management
User-Generated Content following CRUD principles
Production Server Deployment with Gunicorn and Git Version Control

DnD with Toph Exploratory Data Analysis

A financial assessment and report for ‘DnD with Toph’, an online Dungeons & Dragons adventure service, using Python and Pandas to clean, transform, and analyze the data.
Used data to observe growth and profit trends to optimise future scheduling and projects.
Skills include Data Manipulation & Analysis using Python & Pandas and Data Visualization with Matplotlib.

Zombie Survival Game

Developed a Python-based Zombie Survival Game, the mechanics involve managing resources, mission outcomes, and end-of-day round Zombie attacks.
Object-oriented programming in Python to create Class Models for Camp, Locations, and Players.
Utilized smaller functions for specific game mechanics and larger functions to implement and run game flow, mission outcomes, and end-of-round events.
Employed random library and computer decision logic trees for computer-controlled players.

AI / Data Science

Fine-tuning LLM for RPG Statblocks

Created a Fine-tuned task specific model using Llama3 and Unsloth to create Dungeons & Dragons statblocks and game elements.
Deployed on Hugging Face for accesible Training Dataset and LoRA Fine-tuned Adaptors.

Handwritten Character Recognition/Automated Scorecard Processing

Created a Convolutional Neural Network in TensorFlow to predict the class of a handwritten character.
Artificially constructured characters with a strikethrough.
Used measures to avoid overfitting a Computer Vison task, such as Dropout, Batch Normalization, Kernal Regulization, and Data Augmentation.
Achieved a 99% prediction accuracy on the test data.
Model deployed for real world application of Education Scoresheet recogniton.

Smoker vs. Non-Smoker Classification

Created a Neural Network Stacking Classifier using Scikeras KerasClassifier and Scikit-learn’s StackingClassifier.
Used analysis of feature outliers and distrubtion to improve feature engineering.
Combined Catboost, XGBoost, LightGBM and Random Forest Classifier models to create a balanced ensemble model.
Utilized Tensorflow 2.0 Sequential Neural Network for improved ensemble weights.
Achieved a final score of 0.87583 (compared with best of 0.87946), placing 283 out of 1910 submisions.

Red Card / Goal Exploratory Data Analysis

Processed and analyzed data across 20,000 European Football matches, using Pandas for data cleaning, feature engineering and data analysis.
Conducted hypothesis testing with statistical models, including Poisson Means Tests and Linear Regression, to assess the relationship between red cards and goal-scoring.
Visualized distribution patterns and event correlations using Matplotlib and Seaborn.

EIFO Data Extraction/Web Scraping

Built a Python script to extract country risk and cover polcies from EIFO website.
Used BeautifulSoup and Selenium (for dynamic website features) for data selection and extraction.
Cleaned and transformed data in Pandas.
Exported data to Excel output sheet.

Electric Vehicle kWh Consumption Forecasting

Explored forecasting daily kWh consumption for a workplace EV charging program and implemented naive, neural network, and LightGBM prediction models.

Bitcoin Price Forecasting

Implemented time series models (Conv1D, Bidirectional LSTM, GRU) for Bitcoin price forecasting, using metrics (Loss, MAE, MSE) to assess performance.

Structured Medical Text Classification with NLP

Developed a multi-class NLP model utilizing Universal Sentence Encoder, Conv1D character embedding, and text position features to classify medical literature sentences, enhancing readability with structured paragraphs based on classification headings.

Transfer Learning in Food Image Classification

Developed computer vision model on Food101 dataset utilizing EfficientNetV2 transfer learning to achieve over 80% accuracy in categorizing 101 types of food images.

Dungeons & Dragons Race Classification

Utilized Machine Learning techniques with Scikit-learn to predict character races based on ability scores and features from a Dungeons & Dragons dataset.
Explored various models for selection, Random Forest, Logistic Regression, Support Vector Classifier (SVC), K-Nearest Neighbors (KNN), and AdaBoost Classifier.
Conducted in-depth feature analysis and model evaluation.
Fine-tuned hyperparameters of the best performing classfier through manual tuning, RandomizedSearchCV and GridSearchCV.
Achieved an accuracy of 67.25% in predicting character races, improving on 51.8% from the initial dataset.