LA
All projects
AI / MLResearch2026

UltraLLM

Compact language model experiments

Role / Solo — training and interface

UltraLLM — Compact language model experiments
GitHub star
1
curated datasets
3
mobile test UI
Chat

The problem

Most LLM work assumes GPU clusters. I wanted to see how far a small, purpose-built model could go on a constrained budget.

Approach

  • Curated three labelled conversation datasets as YAML, split by intent.
  • Trained a compact model and iterated on tokenization and prompt format.
  • Built an Expo chat client so responses could be judged in a realistic interface.

Result

  • Public repository with dataset, training code and chat client.
  • Findings fed into how I scope AI features for client products.

Stack

  • Python
  • PyTorch
  • Expo
  • React Native
  • YAML datasets