/* Mellon — Data infrastructure for ML (Garuda-style data loader) */ function MellonPage() { const [tab, setTab] = React.useState('overview'); return (
DATA INFRASTRUCTURE DROP-IN

Mellon

Zero-wrangling video data infrastructure for ML. Hardware decode, zero-copy, no pipeline changes.

{/* Hero stats */}
{[ { label: 'Faster training', val: '4.0×', sub: '9.9d → 39.5d, SmolVLM2', accent: true }, { label: 'Lower data prep', val: '4.4×', sub: '7.7s vs 33.6s @ 256²' }, { label: 'Faster inference', val: '4.3×', sub: '207 vs 48 fps @ 1080p' }, ].map((s, i) => (
{s.label}
{s.val}
{s.sub}
))}
{[ { id: 'overview', label: 'Overview' }, { id: 'datasets', label: 'Datasets · 4' }, { id: 'pipelines', label: 'Pipelines · 7' }, ].map(t => ( ))}
{tab === 'overview' && (
Drop-in DataLoader python 
{`# Before — torchvision
from torch.utils.data import DataLoader
loader = DataLoader(VideoDataset(...))

# After — Mellon (no pipeline changes)
from mellon import MellonLoader
loader = MellonLoader(
    dataset="surveillance-jan-2026",
    batch_size=32,
    decode="hardware",      # NVDEC zero-copy
    sample="uniform",
    transform=clip_transform,
)

for batch in loader:
    out = model(batch)      # 4× faster, same API`}
Active pipelines
{[ { name: 'smolvlm2-pretraining', tput: '7.2 GB/s', status: 'streaming' }, { name: 'detection-finetune-yolov8', tput: '4.8 GB/s', status: 'streaming' }, { name: 'caption-eval', tput: '—', status: 'idle' }, ].map((p, i) => (
0 ? '1px solid var(--border-subtle)' : 'none' }}>
{p.name}
{p.tput}
))}
Throughput · last hour
{Array.from({ length: 24 }).map((_, i) => (
19 ? 'var(--accent)' : 'var(--accent-subtle)', borderRadius: 2, }} /> ))}
)} {tab === 'datasets' && (
{['Dataset', 'Hours', 'Samples', 'Format', 'Status', ''].map((h, i) => ( {h} ))}
{[ { name: 'surveillance-jan-2026', hrs: '312.4', samples: '4.2M', format: '.dwarf', status: 'ready' }, { name: 'product-demo-archive', hrs: '24.1', samples: '288K', format: '.mp4', status: 'ready' }, { name: 'dashcam-fleet-q1', hrs: '988.0', samples: '11.2M', format: '.dwarf', status: 'indexing', progress: 0.62 }, { name: 'ugc-moderation-pool', hrs: '1,820', samples: '21.8M', format: '.mp4', status: 'ready' }, ].map((d, i) => (
{d.name} {d.hrs} {d.samples} {d.format}
))}
)} {tab === 'pipelines' && (

Pipeline view coming soon

Visualize your training/inference pipelines and trace decode→model latency end-to-end.

)}
); } /* Anvesha — Agentic search */ function AnveshaPage() { const [query, setQuery] = React.useState('Show me every time a person crossed the perimeter after sunset, then linger near the gate for >30s'); const [running, setRunning] = React.useState(false); const [done, setDone] = React.useState(false); const run = () => { setRunning(true); setDone(false); setTimeout(() => { setRunning(false); setDone(true); }, 1800); }; const trace = [ { step: 1, agent: 'Planner', action: 'Decompose query into 3 sub-tasks', detail: 'detect persons → temporal filter → spatial dwell-time', t: '0.4s' }, { step: 2, agent: 'Retriever', action: 'Search Marengo embeddings', detail: 'index: surveillance-jan-2026 · 1,284 videos', t: '1.2s' }, { step: 3, agent: 'Filter', action: 'Apply astronomical sunset filter', detail: 'lat/lon: 37.77, -122.42 · jan 2026', t: '0.2s' }, { step: 4, agent: 'Reasoner', action: 'Compute dwell time near gate region', detail: 'bbox tracking + IoU > 0.4', t: '3.1s' }, { step: 5, agent: 'Verifier', action: 'Cross-reference with door-sensor logs', detail: 'no entry events — strong signal', t: '0.9s' }, ]; return (
APPLICATION INFRASTRUCTURE AGENTIC

Anvesha

Agentic video search and curation with multi-hop reasoning. Bring your queries — get a bespoke agent.

{[ { label: 'R@1 BodycamBench', val: '50%', sub: 'vs 19% baseline', accent: true }, { label: 'R@1 agentic queries', val: '61%', sub: 'vs 9% baseline', accent: true }, { label: 'Time to custom agent', val: '< 2 min', sub: 'one index, any app' }, ].map((s, i) => (
{s.label}
{s.val}
{s.sub}
))}
{/* Query box */}