Applied AI and machine learning developer building forecasting systems, LLM applications, APIs, and end-to-end data products.

Currently an AI Training & Evaluation Contractor at Mercor Intelligence and an MSc Artificial Intelligence student at the University of Liverpool.

Wilfrid Laurier CS & UX graduate and Humber AI & ML postgrad. Previously worked across AI systems, data, analytics, and internal tools at Arrowz, the Ontario Ministry of Education, DPCDSB, and Meter.

• Football Analytics & Forecasting - probabilistic modelling, player evaluation, expected points, and decision systems
• Applied ML Systems - leakage-aware pipelines, backtesting, simulation, optimization, and deployment
• LLM Memory & Agent Systems - persistent memory, retrieval logic, long-term context, and practical AI workflows
• Multimodal AI & VLMs - combining vision and language for useful tools and interfaces
• Model Training & Optimization - fine-tuning, quantization, evaluation, and efficient inference
• Game Development - building games and interactive systems in Unreal Engine 5

Languages:
Python · TypeScript · SQL · Go · Julia · Java · C# · R · Dart

Machine Learning & AI:
PyTorch · Scikit-learn · TensorFlow · Hugging Face · Transformers · Model Training & Evaluation · Probabilistic Modelling · Simulation

AI Systems & LLMs:
LLM Pipelines · Prompt Engineering · Conversational Memory Systems · Model Integration · LangChain · Dify AI

Backend & APIs:
FastAPI · Flask · REST APIs

Cloud & Data:
AWS (EC2, S3, SageMaker) · PostgreSQL · MongoDB · Apache Spark

Frontend & Additional Tools:
React · Vite · Flutter · Tableau · Figma

End-to-end, time-aware Fantasy Premier League forecasting and squad optimization system.

• Builds leakage-controlled historical datasets and rolling-origin backtests
• Estimates team scoring, clean-sheet, match-outcome, and player-minutes probabilities
• Generates player expected points and outcome distributions using 10,000 joint fixture simulations
• Selects a legal squad, starting XI, bench order, captain, and vice-captain using mathematical optimization
• Publishes forecasts through a static React and TypeScript dashboard

Live dashboard: daniel-mehta.github.io/fpl-forecast

Trained a custom 354M-parameter GPT-style language model using PyTorch and Hugging Face, including the training, evaluation, and inference pipeline.

Led an industry-sponsored capstone combining LLMs with MoFlow and DimeNet++ for molecular generation and property prediction.

Built a Transformer-based recommender with personalized feedback and explainable outputs.

Developed a custom Q-learning system for traffic routing and signal optimization across more than 500 simulated scenarios.