Talk Like

A personalized conversational AI that talks the way I do.

Role
Applied NLP research, CU Boulder
When
Jun 2024 to present
Tools
Llama 3, PEFT, LoRA, RAG, Python
Early telephone operators wearing headsets at a switchboard, in black and white
LLM fine-tuning. A photograph of the setting, not the project itself.

The problem

Talk Like is a highly personalized conversational AI that emulates an individual’s speech patterns and conversational style.

How I built it

  • Persona.

    Llama 3 fine-tuned with PEFT and LoRA on nine years of personal chat data, evaluated through A/B testing.

  • CortexCache.

    A RAG framework that stores dynamic information such as schedules, preferences, and social relationships to enable context-aware interactions.

  • Live data.

    Dynamic APIs feed the CortexCache database in real time, for 20+ user profiles and context windows of up to 5,000 tokens.

Results

94.88%

accuracy in reproducing conversational nuances, by A/B testing.

0.93

G-Eval score for contextual accuracy with CortexCache.

0.21

BLEU score, with distinctive sentence structures.