Add LLM fine-tuning pipeline and F1 halo removal projects

- LLM fine-tuning & on-device inference pipeline: LoRA fine-tuned Llama 3.2
  3B, GGUF quantization, llama.cpp evaluation on 106-case held-out set.
  Results: +0.192 recall over generic baseline at ~half the latency. Domain
  intentionally omitted.
- F1 Halo Removal (CSCI 365 final): classical CV mask detection (Sobel-Y
  arch, probe-and-fit keel) + LaMa spatial inpainting vs RAFT temporal
  propagation on 300-frame 60fps visor-cam footage.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-05-11 02:06:50 -04:00
parent 7e3821d0fa
commit 6e63c992a8
+19 -1
View File
@@ -319,7 +319,7 @@ export const experiences: Experience[] = [
type: "professional",
},
{
title: "Former President, Electrical/Mechanical Team Lead",
title: "President, Electrical/Mechanical Team Lead",
organization: "AIChE Chem-E-Car Competition Team, Bucknell University",
location: "Lewisburg, PA",
period: "Jan 2023 May 2026",
@@ -453,6 +453,24 @@ export const relevantCoursework = [
];
export const projects: Project[] = [
{
title: "LLM Fine-tuning & On-Device Inference Pipeline",
description:
"LoRA fine-tuning and GGUF quantization pipeline for on-device LLM inference. Fine-tuned Llama 3.2 3B improved recall by +0.192 over the generic baseline at roughly half the latency.",
longDescription:
"End-to-end pipeline for fine-tuning a small LLM on structured JSON extraction tasks, then deploying the result on-device. Used LoRA adapters on Llama 3.2 3B via MLX (Apple Silicon) and evaluated across four artifacts: generic F16, generic Q4_K_M, fine-tuned F16, and fine-tuned Q4_K_M. The deployable artifact — fine-tuned Q4_K_M GGUF — achieved 0.961 recall and 100% JSON validity at 2.91s average latency on llama.cpp, compared to 0.780 recall and 6.65s for the generic F16 baseline. Quantization after fine-tuning cost only 0.011 recall while cutting latency in half. Evaluation ran on a 106-case held-out set with automated recall, JSON validity, and schema validity scoring.",
tags: ["Python", "PyTorch", "LoRA", "llama.cpp", "GGUF", "MLX", "Jupyter"],
featured: true,
},
{
title: "F1 Halo Removal via Video Inpainting",
description:
"Removes the F1 Halo safety arch from onboard visor-cam footage using classical CV mask detection and neural inpainting. CSCI 365 final project.",
longDescription:
"The Halo is a mandatory titanium arch on all F1 cars. It saves lives but cuts through the most interesting part of onboard footage. This project removes it cleanly from visor-cam video using two stages. Stage one: classical CV mask detection — Sobel-Y gradient detection for the arch edge, a robust probe-and-fit keel detector with outlier rejection and temporal jump guards, and explicit geometry construction to avoid over-masking. Stage two: two inpainting methods compared side by side — LaMa (Fast Fourier Convolution network, per-frame spatial inpainting) and RAFT optical flow with backward warp and distance-transform blending for temporal coherence across 300 frames at 60fps.",
tags: ["Python", "OpenCV", "Computer Vision", "LaMa", "RAFT", "Jupyter", "Inpainting"],
featured: true,
},
{
title: "Nand2Tetris Implementation (ECEG 431)",
description: