// B05 v4 - Three-tab structure: 'Professional' (resume, default) + 'Personal' (project cards) + 'Academic' (Rice coursework). // Header has name + contact (email + LinkedIn). Tabs swap via JS state, no reload. // Card animations preserved from v3. const B05v3_CardGrid = () => { const TABS = [ { id: 'professional', label: 'Professional', color: 'var(--tab-professional)', fg: 'var(--on-tab-light)' }, { id: 'personal', label: 'Personal', color: 'var(--tab-personal)', fg: 'var(--on-tab-light)' }, { id: 'academic', label: 'Academic', color: 'var(--tab-academic)', fg: 'var(--on-tab-light)' }, ]; const tabFromHash = () => { const h = ((typeof window !== 'undefined' && window.location && window.location.hash) || '').replace('#', ''); return TABS.some((t) => t.id === h) ? h : 'personal'; }; const [tab, setTab] = React.useState(tabFromHash); // deep-linkable: /#personal (default) · /#professional · /#academic const [modalImg, setModalImg] = React.useState(null); React.useEffect(() => { if (!modalImg) return; document.body.style.overflow = 'hidden'; const onEsc = (e) => { if (e.key === 'Escape') setModalImg(null); }; document.addEventListener('keydown', onEsc); return () => { document.body.style.overflow = ''; document.removeEventListener('keydown', onEsc); }; }, [modalImg]); React.useEffect(() => { const expectedHash = '#' + tabFromHash(); if (window.location.hash !== expectedHash) history.replaceState(null, '', expectedHash); const sync = () => setTab(tabFromHash()); window.addEventListener('hashchange', sync); return () => window.removeEventListener('hashchange', sync); }, []); const goTab = (id) => { if (window.location.hash !== '#' + id) history.replaceState(null, '', '#' + id); setTab(id); }; const activeIndex = TABS.findIndex((t) => t.id === tab); const emailTo = () => { window.location.href = 'mailto:' + ['hello', 'andrewmeaux.com'].join('@'); }; // assembled at click; no literal address in source/DOM const css = ` :root{ /* RULE: solid brand hue = active tab only | tinted brand = category tags | tinted green/amber = status pills. White text only on the blue + red active tabs. */ --tab-professional:#2b5aa0; --tab-personal:#b8392c; --tab-academic:#d9a823; --on-tab-light:#ffffff; --tag-engineering-bg:#eaf0f9; --tag-engineering-border:#c8d8ec; --tag-engineering-text:#1f4475; --tag-creative-bg:#fbece9; --tag-creative-border:#efcfc8; --tag-creative-text:#8a291e; --tag-research-bg:#fbf2d8; --tag-research-border:#ecd99a; --tag-research-text:#6e530c; --text-primary:#1c1b16; --text-body:#3a3833; --text-secondary:#5a5a52; --text-muted:#7a7a72; --text-faint:#9a9a92; --text-lightest:#bdbdb5; --bg-page:#f4f3ee; --surface-card:#ffffff; --surface-tint:#ebe9df; --line:#e0ddd0; --line-strong:#ddd9ca; --image-backdrop:#0e1116; --link:#1d4e8e; --status-live-bg:#eaf7ee; --status-live-border:#cae6d4; --status-live-text:#0a7a41; --status-dev-bg:#fbf0e2; --status-dev-border:#ecd9bd; --status-dev-text:#9a5310; --shadow-card-hover:rgba(0,0,0,0.22); } .b05v3{background:var(--bg-page);color:var(--text-primary);font-family:'Inter',system-ui,sans-serif;min-height:100%;padding:64px 56px 72px;font-size:15px;line-height:1.55} .b05v3 a{color:var(--text-primary);text-decoration:underline;text-decoration-color:var(--text-lightest);text-underline-offset:3px} .b05v3 a:hover{text-decoration-color:var(--text-primary)} .b05v3 button,.b05v3 a,.b05v3 summary{-webkit-tap-highlight-color:transparent;touch-action:manipulation} .b05v3 .top{display:grid;grid-template-columns:1fr auto 1fr;align-items:center;margin:0 auto 56px;gap:24px;max-width:1040px} .b05v3 .top .left{display:flex;flex-direction:column;gap:7px;min-width:0} .b05v3 .top .name{font-family:'Inter Tight',sans-serif;font-weight:600;font-size:28px;letter-spacing:-0.015em;color:var(--text-primary);white-space:nowrap;margin:0} .b05v3 .top nav{display:flex;gap:0;justify-self:center;background:var(--surface-tint);border:1px solid var(--line-strong);border-radius:99px;padding:3px;position:relative} .b05v3 .top nav .slider{position:absolute;top:3px;left:3px;bottom:3px;width:calc((100% - 6px)/3);border-radius:99px;z-index:0;transition:transform .32s cubic-bezier(.4,0,.2,1),background-color .32s ease} .b05v3 .top nav button{font-family:'Inter',sans-serif;font-size:13px;letter-spacing:-0.005em;color:var(--text-secondary);background:transparent;border:0;padding:6px 16px;min-width:8em;border-radius:99px;cursor:pointer;font-weight:500;flex:1;position:relative;z-index:1;white-space:nowrap;transition:color .25s} .b05v3 .top nav button:hover{color:var(--text-primary)} .b05v3 .top nav button.on{font-weight:600} .b05v3 .top .contact{display:flex;justify-content:flex-start;gap:18px;font-family:'JetBrains Mono',monospace;font-size:12px;letter-spacing:.01em;flex-wrap:wrap} .b05v3 .top .contact a,.b05v3 .top .contact button{text-decoration:none;color:var(--text-body);display:inline-flex;align-items:center;gap:5px;border:0;border-bottom:1px solid transparent;padding:0 0 1px 0;background:transparent;font:inherit;cursor:pointer;transition:border-color .15s} .b05v3 .top .contact a:hover,.b05v3 .top .contact button:hover{border-bottom-color:var(--text-primary)} .b05v3 .top .contact a .ic{width:11px;height:11px;display:inline-block;flex-shrink:0;opacity:.7} .b05v3 .skip-link{position:absolute;left:-9999px;top:0;background:var(--surface-card);padding:8px 16px;text-decoration:none;color:var(--text-primary);border:1px solid var(--line-strong);border-radius:4px;z-index:100;font-family:'Inter',system-ui,sans-serif;font-size:13px} .b05v3 .skip-link:focus{left:16px;top:16px} .b05v3 .top nav button:focus-visible,.b05v3 .card:focus-visible,.b05v3 .top .contact a:focus-visible,.b05v3 .top .contact button:focus-visible{outline:2px solid var(--link);outline-offset:2px} .b05v3 h2{font-family:'JetBrains Mono',monospace;font-size:11px;letter-spacing:.18em;text-transform:uppercase;color:var(--text-muted);margin:0 0 20px;display:flex;justify-content:space-between;align-items:center} .b05v3 h2 .ct{color:var(--text-lightest)} .b05v3 .grid{display:grid;grid-template-columns:repeat(2,1fr);gap:18px;margin-bottom:56px} .b05v3 .card{background:var(--surface-card);border:1px solid var(--line);border-radius:6px;overflow:hidden;display:flex;flex-direction:column;transition:transform .25s,box-shadow .25s;text-decoration:none;color:inherit;cursor:pointer;font:inherit;text-align:left;padding:0;width:100%} .b05v3 .card:hover{transform:translateY(-3px);box-shadow:0 16px 34px -18px var(--shadow-card-hover)} .b05v3 .card.dead{cursor:zoom-in} .b05v3 .modal-overlay{position:fixed;inset:0;background:rgba(0,0,0,0.88);display:flex;align-items:center;justify-content:center;z-index:1000;padding:40px;cursor:zoom-out;animation:b05v3-fadein .18s ease-out} .b05v3 .modal-img{max-width:100%;max-height:100%;object-fit:contain;box-shadow:0 12px 48px rgba(0,0,0,0.6);border-radius:4px} .b05v3 .modal-close{position:absolute;top:18px;right:22px;background:transparent;border:0;color:#fff;font-size:32px;cursor:pointer;padding:8px 12px;line-height:1;opacity:.85;transition:opacity .15s} .b05v3 .modal-close:hover{opacity:1} @keyframes b05v3-fadein{from{opacity:0}to{opacity:1}} .b05v3 .card .vis{aspect-ratio:5/3;position:relative;overflow:hidden;background:var(--image-backdrop)} /* pause all card animations by default; resume on hover so only the hovered card runs */ .b05v3 .card .vis *,.b05v3 .card .vis *::before,.b05v3 .card .vis *::after{animation-play-state:paused !important} .b05v3 .card:hover .vis *,.b05v3 .card:hover .vis *::before,.b05v3 .card:hover .vis *::after{animation-play-state:running !important} .b05v3 .card .body{padding:18px 20px 22px;display:flex;flex-direction:column;flex:1} .b05v3 .card h3{font-family:'Inter Tight',sans-serif;font-weight:600;font-size:18px;letter-spacing:-0.01em;margin:0 0 4px;display:flex;align-items:center;gap:8px;flex-wrap:wrap} .b05v3 .card .tag{font-family:'JetBrains Mono',monospace;font-size:9.5px;letter-spacing:.12em;text-transform:uppercase;color:var(--text-muted);background:var(--surface-tint);border:1px solid var(--line);padding:2px 7px;border-radius:99px;font-weight:500;white-space:nowrap} .b05v3 .card .tag.research{color:var(--tag-research-text);background:var(--tag-research-bg);border-color:var(--tag-research-border)} .b05v3 .card .tag.creative{color:var(--tag-creative-text);background:var(--tag-creative-bg);border-color:var(--tag-creative-border)} .b05v3 .card .tag.engineering{color:var(--tag-engineering-text);background:var(--tag-engineering-bg);border-color:var(--tag-engineering-border)} .b05v3 .card .tag.live{color:var(--status-live-text);background:var(--status-live-bg);border-color:var(--status-live-border)} .b05v3 .card .tag.dev{color:var(--status-dev-text);background:var(--status-dev-bg);border-color:var(--status-dev-border)} .b05v3 .card .tag.live::before,.b05v3 .card .tag.dev::before{content:"";display:inline-block;width:5px;height:5px;border-radius:50%;background:currentColor;margin-right:5px;vertical-align:middle} .b05v3 .card .linkrow{display:flex;align-items:center;gap:8px;flex-wrap:wrap} .b05v3 .card .meta{font-family:'JetBrains Mono',monospace;font-size:11px;letter-spacing:.04em;color:var(--text-muted);margin-bottom:10px} .b05v3 .card p{font-size:14px;line-height:1.55;color:var(--text-body);margin:0 0 14px;flex:1} .b05v3 .card .arrow{font-family:'JetBrains Mono',monospace;font-size:12px;color:var(--link);letter-spacing:.02em} .b05v3 .card.dead .arrow{color:var(--text-muted)} .b05v3 .card .built{font-family:'JetBrains Mono',monospace;font-size:11px;color:var(--text-faint);letter-spacing:.02em;margin-top:6px} /* ============== RESUME TAB ============== */ .b05v3 .resume{max-width:1040px;margin:0 auto;padding-top:8px} .b05v3 .resume .rh{font-family:'Inter Tight',sans-serif;font-weight:500;font-size:34px;line-height:1.15;letter-spacing:-0.025em;color:var(--text-primary);margin:0 0 8px} .b05v3 .resume .rsub{font-size:15px;color:var(--text-body);margin:0 0 36px;line-height:1.55;min-height:3.1em} .b05v3 .resume section{margin-bottom:38px} .b05v3 .resume section > h3{font-family:'JetBrains Mono',monospace;font-size:11px;letter-spacing:.2em;text-transform:uppercase;color:var(--text-muted);margin:0 0 18px;padding-bottom:8px;border-bottom:1px solid var(--line);font-weight:500} .b05v3 .resume .job{display:grid;grid-template-columns:120px 1fr;gap:22px;padding:14px 0;border-bottom:1px dotted var(--line)} .b05v3 .resume .job:last-child{border-bottom:0} .b05v3 .resume .job .when{font-family:'JetBrains Mono',monospace;font-size:11.5px;letter-spacing:.04em;color:var(--text-muted);line-height:1.5;padding-top:3px} .b05v3 .resume .job .what .title{font-family:'Inter Tight',sans-serif;font-weight:600;font-size:16px;letter-spacing:-0.01em;color:var(--text-primary);margin:0 0 2px} .b05v3 .resume .job .what .org{font-family:'JetBrains Mono',monospace;font-size:12px;letter-spacing:.04em;color:var(--text-muted);margin:0 0 8px} .b05v3 .resume .job .what p{font-size:14px;line-height:1.6;color:var(--text-body);margin:0 0 6px} .b05v3 .resume .job .what ul{margin:6px 0 0;padding:0;list-style:none} .b05v3 .resume .job .what ul li{font-size:14px;line-height:1.55;color:var(--text-body);padding:3px 0 3px 16px;position:relative} .b05v3 .resume .job .what ul li::before{content:"";position:absolute;left:2px;top:11px;width:6px;height:1px;background:var(--text-lightest)} .b05v3 .resume .edu{display:grid;grid-template-columns:120px 1fr;gap:22px;padding:14px 0;border-bottom:1px dotted var(--line)} .b05v3 .resume .edu:last-child{border-bottom:0} .b05v3 .resume .edu .when{font-family:'JetBrains Mono',monospace;font-size:11.5px;letter-spacing:.04em;color:var(--text-muted);padding-top:3px} .b05v3 .resume .edu .what .title{font-family:'Inter Tight',sans-serif;font-weight:600;font-size:16px;letter-spacing:-0.01em;color:var(--text-primary);margin:0 0 2px} .b05v3 .resume .edu .what .org{font-family:'JetBrains Mono',monospace;font-size:12px;letter-spacing:.04em;color:var(--text-muted);margin:0 0 8px} .b05v3 .resume .edu .what .note{font-size:14px;line-height:1.6;color:var(--text-body);margin:0} .b05v3 .resume .what .team{font-family:'JetBrains Mono',monospace;font-size:11px;color:var(--text-faint);letter-spacing:.04em;margin:0 0 8px} .b05v3 .resume .what .paper{font-family:'JetBrains Mono',monospace;font-size:12px;color:var(--link);letter-spacing:.02em;text-decoration:none;display:inline-block;margin-top:2px} .b05v3 .resume .what .paper:hover{text-decoration:underline} .b05v3 .resume .what .chips{display:flex;flex-wrap:wrap;gap:6px;margin:2px 0 12px} .b05v3 .resume .what .chip{font-family:'JetBrains Mono',monospace;font-size:10px;letter-spacing:.02em;color:var(--text-secondary);background:var(--surface-tint);border:1px solid var(--line);border-radius:99px;padding:2px 8px;white-space:nowrap} .b05v3 .resume .what details.case{margin:0 0 12px} .b05v3 .resume .what details.case > summary{font-family:'JetBrains Mono',monospace;font-size:10.5px;letter-spacing:.08em;text-transform:uppercase;color:var(--link);cursor:pointer;list-style:none;display:inline-block;padding:2px 0} .b05v3 .resume .what details.case > summary::-webkit-details-marker{display:none} .b05v3 .resume .what details.case > summary::before{content:"▸ ";color:var(--text-muted)} .b05v3 .resume .what details.case[open] > summary::before{content:"▾ "} .b05v3 .resume .what .cs{font-size:13px;line-height:1.55;color:var(--text-body);margin:7px 0 0;max-width:760px} .b05v3 .resume .what .cs .csk{font-family:'JetBrains Mono',monospace;font-size:10px;letter-spacing:.05em;text-transform:uppercase;color:var(--text-muted);display:block;margin-bottom:2px} .b05v3 .resume .what .pipe{font-family:'JetBrains Mono',monospace;font-size:10.5px;letter-spacing:.01em;color:var(--text-muted);margin:0 0 12px;line-height:1.5} /* projects tab small intro */ .b05v3 .pintro{font-size:15px;color:var(--text-body);margin:0 0 28px;line-height:1.55;max-width:680px} /* ============== RESPONSIVE ============== */ /* Small laptop / tablet landscape — drop card grid to 2 columns, ease padding */ @media (max-width: 980px) { .b05v3{padding:48px 32px 56px} .b05v3 .grid{grid-template-columns:repeat(2,1fr);gap:16px} } /* Tablet portrait / large mobile — stack header, 1-col cards, stack resume entries */ @media (max-width: 720px) { .b05v3{padding:28px 18px 40px;font-size:14px} .b05v3 .top{grid-template-columns:1fr;gap:14px;margin-bottom:32px} .b05v3 .top nav{justify-self:stretch;width:100%} .b05v3 .top nav button{flex:1;padding:8px 10px} .b05v3 .top .contact{justify-content:flex-start;gap:14px;flex-wrap:wrap} .b05v3 .grid{grid-template-columns:1fr;gap:14px;margin-bottom:36px} .b05v3 .card h3{font-size:17px} .b05v3 h2{margin-bottom:14px} .b05v3 .pintro{margin:-2px 0 18px} .b05v3 .resume{padding-top:0} .b05v3 .resume .rh{font-size:26px} .b05v3 .resume .rsub{font-size:14px;margin-bottom:24px} .b05v3 .resume section{margin-bottom:28px} .b05v3 .resume section > h3{margin-bottom:12px} .b05v3 .resume .job, .b05v3 .resume .edu{grid-template-columns:1fr;gap:4px;padding:14px 0} .b05v3 .resume .job .when, .b05v3 .resume .edu .when{padding-top:0;font-size:11px} .b05v3 .resume .job .what .title, .b05v3 .resume .edu .what .title{font-size:15px} .b05v3 .resume .job .what p, .b05v3 .resume .job .what ul li{font-size:13.5px} } /* Small mobile — tighter padding + scaled type */ @media (max-width: 420px) { .b05v3{padding:22px 14px 32px} .b05v3 .top .name{font-size:22px} .b05v3 .top .contact{font-size:11.5px;gap:12px} .b05v3 .top nav button{font-size:12px;padding:7px 8px} .b05v3 .resume .rh{font-size:23px} } `; return (
Skip to content

Andrew Meaux

LinkedIn ↗ GitHub ↗
{tab === 'professional' &&

Enterprise AI engineer designing industrial agentic systems, with 11 years across offshore operations & engineering, digital portfolio ownership, and enterprise AI delivery, including competitive selection for Chevron’s Digital Scholar Program and completion of a full-time Professional Master of Data Science (MDS) with specialization in machine learning at Rice University.

Experience

Nov 2025 - Present
AI Engineer - Enterprise AI
Chevron · Houston, TX

Translating endorsed agentic AI strategies and emerging technologies into production-oriented architectures, agent patterns, and working implementations. Lead engineer on a multi-agent deep research system that automates long-horizon research across unified external and internal data streams. Engaging directly with technical and business teams to translate applied AI between technical depth and business value.

Nov 2024 - Nov 2025
Systems Engineer - Enterprise AI
Chevron · Houston, TX

Acted as solutions architect for the enterprise's exploration business, translating business needs and value drivers into target AI capabilities and a delivery roadmap. Co-authored an internal "Perspective on Agentic AI" whitepaper establishing a practical AI agent taxonomy and governance pattern that shaped cross-functional partnerships and foundational agentic AI and data strategies.

Sep 2023 - Oct 2024
Digital Integration Lead - Offshore Wells
Chevron · Houston, TX

Sole business and technology liaison to the enterprise digital platform and cross-functional partners, with full ownership of the BU's digital strategy, portfolio, and demands.

Aug 2022 - Aug 2023
Digital Scholar Program
Chevron · Rice University · Houston, TX

Selected for competitive, company-sponsored educational program. Completed full-time, one-year Professional Master of Data Science (MDS) with specialization in Machine Learning at Rice University.

Mar 2019 - Jul 2022
Lead Field Engineer - Offshore Wells
Chevron · Deepwater Gulf of Mexico

Onsite lead engineer (team of 3) on deepwater drilling & completions platform managing high-stakes, complex operations at ~$1M/day spread-rate in support of safe, incident-free operations, engineering execution, and offshore logistics.

Sep 2017 - Mar 2019
Field Engineer - Offshore Wells
Chevron · Deepwater Gulf of Mexico
May 2016 - Sep 2017
Accelerated Drilling Program - DWEP
Chevron · Deepwater Gulf of Mexico
Jun 2015 - May 2016
Drill / Well Site Manager - Offshore East
Chevron · Gulf of Mexico Shelf

Education

2022 - 2023
Professional Master of Data Science (MDS)
Rice University · Houston, TX
2011 - 2015
B.S., Petroleum Engineering
Louisiana State University · Baton Rouge, LA

Publications

2024
Editor & Self-Publisher - Posthumous Poetry Collection
Family Archive Project
Edited and published a curated posthumous collection of my late grandfather's poetry, preserving South Louisiana cultural history. Managed manuscript prep, book design, ISBN/metadata, and print-on-demand.
} {tab === 'personal' &&

Selected independent builds exploring agentic software, retro interfaces, AI-native games, and model evaluation. These projects are built outside work through agentic AI engineering workflows and shown through screenshots, live links, and short writeups. They’re how I test evolving agent capabilities, workflows, and features upon release.

Projects

{/* ============ PEGPOINT - only live link ============ */}
Pegpoint screenshot

Pegpoint CREATIVE

retro gaming · pixel art · procedural audio

A retro pixel-art puzzle game with visuals, audio, physics, and levels programmatically generated by Claude Design and Claude Code. Built for Vibe Jam 2026, a video-game hackathon.

pegpoint.io ↗live
built with Claude Design x Claude Code
 {/* ============ RUFF & HONOURS - in development ============ */} {/* ============ AGENT HUB '98 - private, dead link ============ */} {/* ============ ESUNA - in development (hidden from prod) ============ */} {false && (
Esuna screenshot

Esuna RESEARCH

typed language · bounded narration · replayable

A tiny experimental typed language for bounded LLM tabletop RPGs. Code owns the rules, dice, and state, while the LLM owns bounded narration. Every prompt has a return type, every action must compile, making the seam between deterministic mechanics and model improvisation explicit, typed, replayable, and measurable.

esuna.orgin development
built with Claude Code
)} {/* ============ CYBULATION - in development ============ */} {/* ============ ROUGAROUX - in development (hidden from prod) ============ */} {false && (
Rougaroux screenshot

Rougaroux RESEARCH

safety · interpretability · social-deduction

A pre-registered interpretability study: can game-defined deception be read from an open-weight model's activations as LLM agents play a Werewolf-like social-deduction game? Ships with an observable replay: agent view, omniscient view, and a live probe-vs-ground-truth overlay. Game-defined deception, not a lie detector.

rougaroux.orgin development
built with Claude Code
)} {/* ============ AGENTS ONLINE - in development, dead link (hidden from prod) ============ */} {false && ( )}
} {tab === 'academic' &&

Selected group projects from my full-time Professional Master of Data Science (MDS) with specialization in machine learning at Rice University (2022–2023), completed as part of Chevron’s Digital Scholar Program. Each summary reflects the final team report, with links to the source paper and technical details drawn from it.

Projects

Summer 2023
EnGenIR — RAG for engineering Q&A
retrieval-augmented LLM · vector embeddings · prompt engineering
D2K Capstone · team of 6

Retrieval-augmented question-answering over facilities-engineering pipeline standards: it chunks and embeds technical specifications, retrieves the relevant passages, and answers design questions through a prompt-engineered LLM with cited source context.

84.0% recall@10250K+ words indexed23-question blind eval3 standards + synthetic spec
Standards → chunks → embeddings → retrieval → Power Prompt → cited answer
Details
ProblemEngineering standards are long, technical, and high-consequence; answering a design question often means synthesizing requirements across several documents.
ApproachParse and chunk the standards, embed the passages, retrieve the relevant context, and combine it with the query in an engineered "Power Prompt."
ResultE5-Large-v2 reached 84.0% recall@10; in a 23-question project evaluation, EnGenIR outscored zero-shot Claude 2 and GPT-4 on relevance and accuracy.
LimitationFocused on textual requirements; tables, equations, and semantic chunking are future work.
TakeawayIn high-consequence domains, retrieval quality and prompt discipline are what make LLM answers usable and grounded.
Read the paper ↗
Summer 2023
HARP — human vs. AI music detector
mel-spectrogram classifier · HPSS · ViT-L/16 · PyTorch
group project · team of 3

Binary classifier that flags AI-generated music from raw audio. Audio is converted to mel-spectrograms and split into harmonic and percussive components (HPSS), then classified with a Vision Transformer.

81.84% test accuracy0.901 AUROCHPSS + ViT-L/16MusicLM unseen in test
Raw audio → mel-spectrogram → HPSS → harmonic / percussive → ViT → human / AI
Details
ProblemAs AI-generated music gets easier to produce, platforms and artists need to identify model-generated recordings.
ApproachNormalize audio, convert to mel-spectrograms, separate harmonic and percussive components (HPSS), and classify with a Vision Transformer.
Result81.84% test accuracy at 0.901 AUROC, with an entire generator (MusicLM) held out for the test set.
LimitationGeneralization depends on dataset diversity; newer music models would need broader retraining data.
TakeawayTreating audio as spectrogram images let an image-classification architecture do AI-music detection.
Read the paper ↗
Spring 2023
VIBE — album artwork to music
CLIP · GPT · Jukebox · multimodal
group project · team of 4

Multimodal pipeline that turns album artwork into genre-conditioned music: CLIP predicts genre from the cover, an LLM writes lyrics, and Jukebox is conditioned on artist, genre, and lyrics to generate audio.

56.3% top-178.6% top-3136.5K mapped albumsbest vs. 4 human testers
Album art → CLIP genre → GPT lyrics → Jukebox → music
Details
ProblemImage-to-music chains vision, language, and audio models across noisy intermediate representations.
ApproachBuild a Spotify album-art dataset (143,109 images, 136,507 mapped to 15 parent genres), context-optimize CLIP, generate lyrics, and condition Jukebox on genre, artist, and lyrics.
ResultContext optimization lifted CLIP genre top-1 from 30.9% to 56.3% (78.6% top-3); on a 75-image blind set, optimized VIBE-CLIP ranked first, beating all four human testers.
LimitationJukebox was computationally expensive, so generation was limited and audio quality leaned heavily on artist conditioning.
TakeawayMultimodal chains are compelling, but every intermediate prediction is a new failure point.
Read the paper ↗
Spring 2023
GraMR — graph-based music recommendations
GraphSAGE · GNN · PyTorch Geometric
group project · team of 2

Graph neural network for playlist generation from a seed song: songs and playlists form a bipartite graph, GraphSAGE learns embeddings, and link prediction over them generates coherent playlists.

0.93 link-prediction AUC100K playlists681,660 song nodes6.59M edges
Spotify playlists → bipartite graph → GraphSAGE → link prediction → playlist
Details
ProblemPlaylist recommendation blends collaborative filtering, content features, and subjective musical cohesion.
ApproachConvert Spotify playlist data into a song-and-playlist bipartite graph and train a GraphSAGE link-prediction model.
ResultLink-prediction AUC scaled with data: 0.68 at 1K playlists, 0.83 at 10K, and 0.93 at 100K.
LimitationUsed one-tenth of the full Million Playlist Dataset; playlist quality was judged subjectively.
TakeawayScale mattered: average song-node degree rose from 1.9 to 9.7 as playlists grew, improving link-prediction performance.
Read the paper ↗
}
{modalImg && (
setModalImg(null)}> {modalImg.alt} e.stopPropagation()} />
)}
); };