Developed an end-to-end serverless fine-tuning pipeline to adapt Meta Llama 3.1 8B Instruct for hyper-personalized conversational style and Hinglish tone replication. Architected multi-platform data harvesters to extract, sanitize, and format 10,000+ Signal and Instagram message logs into dynamic 12-turn dialogue windows. Trained a 4-bit QLoRA adapter using Unsloth and Hugging Face SFTTrainer on serverless NVIDIA A100 GPUs via Modal Cloud.
Prompt: "what does love look like to you?"
Response: "Well... You put your phone aside for a while. That's love to me."