I am a Software Engineering student at the University of California, Irvine interested in innovative applications of AI.

Speculative decoding across the Apple Neural Engine and GPU on a single chip, using public APIs only.

• Stack: Swift, Python, Core ML, MLX/Metal.
• Key Feature: Runs a draft model on the ANE and a target model on the GPU in parallel, achieving a 1.47× speedup over GPU-only decoding with the ANE draft fully hidden inside the GPU's verify window. Includes a custom measurement harness and handoff characterization protocol.

A local development environment and orchestration layer for open-source LLMs.

• Stack: Python, FastAPI, Docker, Ollama / Hugging Face.
• Key Feature: Provides an asynchronous API wrapper to deploy and run quantized open-source models locally with optimized inference and custom agent workflows.

A web app that turns rare-earth-element solvent-extraction research papers into a structured, queryable database.

• Stack: Python, Anthropic API (Claude Opus/Haiku), Streamlit, SQLite, pdfplumber.
• Key Feature: Digitizes figures from research PDFs via a versioned Claude extraction pipeline, grounds the model against a deterministic vector-geometry marker count to catch under-digitization, and gates every result through an automatic QA suite plus human review before it's merged — with a strict read/write boundary so the query-facing AI assistant can't mutate data even if jailbroken.

A LoRA fine-tuning pipeline for a Python code-generation LLM, rebuilt end-to-end for Apple Silicon.

• Stack: Python, MLX, Hugging Face, Anthropic API.
• Key Feature: Fine-tuned Qwen2.5-Coder-3B on 28K synthetic OSS-Instruct pairs, lifting HumanEval pass@1 by 3.0 pts (111→116/164) and MBPP by 1.4 pts — ported the original CUDA-only QLoRA recipe to mlx-lm and 4-bit quantization, then published the tuned adapter to Hugging Face Hub.

• LinkedIn: linkedin.com/in/nicholas-tweedie-425a9b32b
• Email: ntweedie@uci.edu · nttweedie@gmail.com