Cultural memory infrastructure for New York City, running entirely on local hardware.

A 3D time machine for every building in NYC. 1,082,831 structures across all five boroughs, extruded to real LiDAR roof heights, with a year scrubber from 1700 to 2026. Click any building and a 30B language model generates a biography from public NYC datasets. No cloud, no internet at runtime.

Ghost buildings remember what the city demolished: the Twin Towers rise in 1970 and vanish in 2001, Penn Station disappears in 1963, 117 Cross-Bronx tenements collapse year by year as Robert Moses displaces 60,000 people. 750 archival photos from the NYPL Milstein Division are pinned to the map. Immigration flows animate across centuries.

Two renderers consuming one shared data spine (building polygons, landmarks, ghost manifests, cultural content).

Native renderer (this branch) -- Bevy 0.18.1 + wgpu in Rust. Real 3D scene with directional lighting and shadows. Demolished buildings loaded as artist-authored .glb meshes via an ECS manifest system (each ghost is a Bevy entity with coordinate, era window, and niche tags). PS2-era post-process as GPU shaders. Shares 128 GB unified memory with a 30B language model on the DGX Spark. Includes a FastAPI orchestrator for biography generation via NVIDIA Nemotron-3 Nano 30B.

Browser demo (feature/sketch-overlay) -- deck.gl + MapLibre in a single index.html. Extruded polygons on a flat map, 8 era color bands, immigration particle simulation, NYPL photo pins, Sobel sketch overlay. Runs on any laptop with a browser.

# Browser demo (any machine)
git checkout feature/sketch-overlay
python3 -m http.server 8080
# open http://localhost:8080/index.html

# Native renderer + orchestrator (requires Rust toolchain)
cd renderer-rust && cargo run

# Orchestrator in stub mode (any laptop, no GPU)
pip install -r requirements.txt
./scripts/dev-stub.sh
# live on http://localhost:30000

# On the DGX Spark (real inference)
./scripts/start-llama-nano.sh
NARRATION_MODE=real LLAMA_SERVER_URL=http://127.0.0.1:8090 \
  python -m uvicorn src.orchestrator.main:app --host 0.0.0.0 --port 30001

• Bevy 0.18.1 + wgpu (native renderer, Rust)
• deck.gl 9.1.4 + MapLibre GL 3.6.2 (browser demo)
• NVIDIA Nemotron-3 Nano 30B via llama.cpp
• FastAPI orchestrator (biography RAG, caching)
• Acer GN100 / NVIDIA DGX Spark / GB10 Grace Blackwell

Built at Spark Hack Series NYC, Apr 2026. Cultural Impact track winner + Most Likely to Become a Unicorn bounty.

Special thanks to Carson Weeks (Rust renderer), James Burke (cultural content), and Marvens Destine (hackathon networking).