NYC Smart City Nexus

Energy, Waste & Urban Intelligence Platform

Spark Hack NYC 2026 | Environmental Impact Track | NVIDIA GB10

A decision-support platform that merges NYC's energy and waste data into one unified intelligence layer. Built to run entirely on-device using NVIDIA RAPIDS, NIMs, and cuOpt on the Acer Veriton GN100.

---

Quick Start

# Run the full pipeline
python smart-city-management/build_silver_layer.py
python smart-city-management/build_bridge_layer.py
python smart-city-management/build_spatial_layer.py
python smart-city-management/build_gold_layer.py
python smart-city-management/add_missing_outputs.py

---

Project Structure

smart-city-management/
├── build_silver_layer.py     # Raw → Silver (12 Parquet datasets)
├── build_bridge_layer.py    # Create property_to_bbl bridge
├── build_spatial_layer.py  # cuSpatial feature joins
├── build_gold_layer.py     # → Gold (3 tables + handoffs)
├── add_missing_outputs.py   # depots + site_profiles
├── README.md
├── design doc/              # Technical design docs
├── frontend/                # Dashboard UI
├── AI/                     # LLM ranker
└── data/
    ├── raw/                # 12 raw CSVs from NYC Open Data
    ├── silver/              # Cleaned Parquet (partitioned)
    ├── gold/                # Unified tables + handoffs
    │   ├── route_inputs/    # cuOpt inputs
    │   ├── dispatch/        # BESS dispatch
    │   └── nim/            # NIM site batches
    └── dictionaries/        # Data documentation

---

Datasets (12 sources from NYC Open Data)

Energy

Status	ID	Dataset	Rows	Source
✅	E1	Energy Cost Savings Program	2,363	NYC Open Data
✅	E3	Electric Consumption & Cost	553,666	NYC Open Data
✅	E4	NYC EV Fleet Station Network	1,638	NYC Open Data
✅	E5	Municipal Solar Readiness (LL24)	4,270	NYC Open Data
✅	E7	LL84 Monthly Energy Data	~2.2M	NYC Open Data
✅	E10	LL84 Annual Benchmarking	103,259	NYC Open Data

Waste

Status	ID	Dataset	Rows	Source
✅	W1	DSNY Monthly Tonnage	24,883	NYC Open Data
✅	W2	311 Service Requests (DSNY)	1,556,494	NYC Open Data
✅	W3	Litter Basket Inventory	20,413	NYC Open Data
✅	W7	Food Scrap Drop-Off Locations	591	NYC Open Data
✅	W8	Disposal Facility Locations	39	NYC Open Data
✅	W12	Waste Characterization	2,112	NYC Open Data

---

Pipeline Outputs

Silver Layer (data/silver/)

File	Description	Partitioned
E3_electric_consumption.parquet	NYCHA electric usage	No
E4_ev_fleet_stations.parquet	EV charging locations	No
E5_solar_readiness.parquet	Solar-ready buildings	No
E5_solar_readiness_enriched.parquet	+ cuSpatial features	No
E7_ll84_monthly.parquet	Monthly energy	✅ calendar_year
E10_ll84_benchmarking.parquet	Annual benchmarking	No
property_to_bbl.parquet	Bridge table (48,310 unique)	No
W1_dsny_monthly_tonnage.parquet	District tonnage	No
W2_311_dsny.parquet	311 complaints	✅ created_year
W7_food_scrap_dropoffs.parquet	Compost sites	No
W8_disposal_facilities.parquet	Transfer facilities	No

Gold Layer (data/gold/)

File	Description
unified_sites.parquet	Gold Table 1: E5 + E10 + spatial features
district_waste.parquet	Gold Table 2: W1 aggregated by district
time_series.parquet	Gold Table 3: E7 monthly energy
route_inputs/depots.parquet	cuOpt depot locations (39)
route_inputs/demand_nodes.parquet	cuOpt demand nodes (59 districts)
nim/site_batches.jsonl	NIM scoring batch (20 sites)
dispatch/site_profiles.parquet	BESS dispatch profiles (4,268 sites)

---

Interface Contracts (GB10 Runbook)

Join Rules

E5.bbl ← E10.bbl (via property_to_bbl)
E7.property_id → E10.property_id → E10.bbl → E5.bbl

cuSpatial Features

From E5 building points:

• EV ports within 500m
• EV ports within 1km
• Compost sites within 1km
• Nearest transfer facility distance

Complaint Schema

Map both Problem/Problem Detail and Complaint Type/Descriptor into one canonical category column.

---

Tech Stack

Layer	Tool	Purpose
Data Processing	RAPIDS cuDF + cuSpatial	GPU-accelerated data processing
Route Optimization	RAPIDS cuOpt	Waste collection vehicle routing
LLM Inference	NVIDIA NIMs (Llama 3.1 8B)	Local site scoring
Simulation	Custom Python (GPU-vectorized)	BESS dispatch
Dashboard	Streamlit + Folium	Interactive maps, charts

---

Running the Next Stages

cuOpt (Route Optimization)

import cuopt

# Load demand nodes and depots
demand_nodes = pd.read_parquet("data/gold/route_inputs/demand_nodes.parquet")
depots = pd.read_parquet("data/gold/route_inputs/depots.parquet")

# Solve route
# (See: https://docs.nvidia.com/cuopt/user-guide/latest/introduction.html)

NIM (Local LLM Scoring)

# Run NIM container
docker run -d --gpus=1 --name nimLlama \
  -v ./smart-city-management/data/gold/nim:/app/nim \
  nvcr.io/nim/nvidia/llama-3.1-nim:8b-instruct-q4

# Score sites
curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d @nim/site_batches.jsonl

BESS Dispatch

# Load site profiles and time series
profiles = pd.read_parquet("data/gold/dispatch/site_profiles.parquet")
time_series = pd.read_parquet("data/gold/time_series.parquet")

# Run dispatch simulation
# (See: design doc/BESS_simulation.md)

Future Things to Implement

Local LLM Scoring (Qwen 3 80B)

The mini workstation already has the Qwen 3 80B model activated and available for local inference. For future local LLM scoring of sites, we can leverage this model directly:

# Using the local Qwen 3 80B endpoint (already running on the workstation)
curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-80b",
    "messages": [{"role": "user", "content": "Score this site for solar potential..."}]
  }'

The Qwen 3 80B provides significantly more capacity than Llama 3.1 8B NIM for nuanced site scoring, natural language reasoning about building characteristics, and batch processing of site evaluations.

---

Team

Built for Spark Hack NYC 2026 — Environmental Impact Track.

NVIDIA GB10: 128 GB unified memory, 1 PFLOP FP4 on Acer Veriton GN100

Dataset & Model Training

Datasets Used

• Garbage Classification Dataset (Kaggle)
• Garbage Detection Dataset (Roboflow)

Data Preparation

Both datasets were preprocessed and reorganized into four unified classification categories:

• Organics
• Electronic
• Plastic
• Others

Images from both sources were partitioned into folder-based class directories to maintain consistent labeling and balanced training structure.

Base Model

• Model Name: NVIDIA Nemotron Nano 12B v2 VL FP8
• Model ID: nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-FP8

Fine-Tuning Method

This project uses LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning, enabling adaptation of the multimodal vision-language model with lower memory and compute overhead.

Training Goal

The objective is to fine-tune the model for garbage image classification across mixed-source datasets, improving recognition accuracy for real-world waste sorting applications.