GPT-4o fine-tuning
for customer
live chat.

Two-dimensional projection of product names, short and full descriptions, and company information embeddings

Behind the visual

A map of the AI’s knowledge.

I implemented the knowledge base in Python as a JSONL dataset, combining product information with company policies and staff training documents.

The below graph shows the position vectors, each with 1536 dimensions, projected onto a two-dimensional plane that preserves 20.6% of the variance.

How I built it

Selected work.

3 projects shown

Frederick Searancke

Frederick Searancke

Quant Researcher Computer Science, University of Warwick

About me

Skills in practice.

Machine learning

  • EBM model training
  • Loss functions & regularisation
  • Overfitting & underfitting diagnostics
  • LLM fine-tuning & RAG

Quantitative research

  • Walk-forward backtesting
  • Hypothesis testing & null models
  • Parameter sweeps & overfitting
  • Volatility & slippage modelling

Software & data

  • Python and Java
  • SQL
  • Data pipelines
  • Feature engineering

Building systems

  • End-to-end project delivery
  • Modular system architecture
  • TCP APIs · IBKR TWS
  • REST API integration