Experience 2026

EY — AI & Data

Domain-adapting open-source LLMs for finance

I built a local pipeline around EDGAR filings, RAG, LoRA, and DPO.

01 · Pipeline: EDGAR → RAG / LoRA → DPO
02 · Model: TinyLlama 1.1B
03 · Compute: Apple Silicon

Role

AI & Data Intern at EY, June 2026, Dublin.

What I built

I used SEC filings from EDGAR to adapt TinyLlama-1.1B to financial questions. The pipeline covered data preparation, synthetic examples, RAG, LoRA fine-tuning, DPO, and evaluation. It ran locally on Apple Silicon with PyTorch, Hugging Face, peft, and trl.

Pipeline

SEC filings (EDGAR)
        |
        v
  Data preparation     raw documents → training examples
        |
   -----+-----
   |         |
   v         v
  RAG      Fine-tuning    adapt the model (one or both)
   |         |
   -----+-----
        |
        v
  Alignment (DPO)      nudge outputs toward better answers
        |
        v
  Evaluation           base vs fine-tuned vs RAG

Technical detail

Fine-tuned via rank-16 LoRA adapters on roughly 0.5% of model parameters.
Aligned outputs to preferences using a custom reward model and DPO.
Built an evaluation harness benchmarking base vs fine-tuned vs RAG models on ROUGE-L, BERTScore, and NLI-based hallucination detection.
RAG eliminated the base model's fabricated financial figures in the evaluation set.

Why finance filings

SEC filings are public, large enough to be useful, and easy to audit. That let me show the whole pipeline without using client documents.