Hackathon Project
ChargePilot: EV Charging Network Optimization for Austin
ChargePilot is a decision-support system that ranks where to deploy EV charging capacity using reproducible geospatial features and graph-based optimization instead of ad hoc site selection.
Hackathon Story
This was my first hackathon, and I built ChargePilot with my triplet brothers and Raghu Subramanian. We had just 2.5 hours to go from idea to working demo, which made it a fun sprint in fast problem framing, collaboration, and hands-on learning.
Hackathon Demo
Demo walkthrough of scenario exploration, map interaction, and recommendation outputs.
Team at the Hackathon
three clones and a bug
- Problem Focus
- Network-level EV charging expansion under fragmented supply and demand signals
- Tech Stack
- Python, GeoPandas, graph analytics, Next.js, Leaflet, API routes
- Primary Inputs
- 684 existing stations, 173 candidate parking entrances, traffic and demographic context
- Scenarios
- Budget-constrained expansion plans for 10, 25, and 50 new sites
Problem Context and Goal
Planning EV infrastructure in Austin is difficult because the relevant signals are split across disconnected systems: station inventory, traffic pressure, parking feasibility, and neighborhood access. Local heuristic decisions often fail to optimize at the network level.
Goal: build an auditable planning workflow that ranks where new charging capacity should be deployed using learned scoring and facility-location style optimization.
Objectives
- •Integrate municipal and infrastructure datasets into one planning graph
- •Score candidate sites with a learned model instead of fixed manual weights
- •Select sites under constrained budgets (10, 25, 50) with facility-location logic
- •Render recommendations in an interactive product UI for technical and non-technical users
- •Generate CSV and GeoJSON artifacts for downstream analysis and visualization
Data and Feature Engineering
We combined public charging, parking, traffic, and neighborhood context into one planning dataset, then generated normalized candidate scores to rank expansion sites.
- •684 existing stations and 173 candidate parking entrances
- •Core signals: traffic pressure, parking feasibility, and charger gap
- •Outputs exported as auditable CSV and GeoJSON artifacts
Pipeline Architecture
The build used a simple split: offline Python for deterministic scoring and optimization, then a Next.js plus Leaflet frontend for fast scenario exploration.
- •ETL and normalization into analysis-ready tables
- •Precomputed recommendations copied into web data assets
- •Interactive map UI for comparing budget scenarios
Modeling and Optimization Design
Learned Node Scoring
- •Ridge regression with lambda = 0.05 learns feature weights
- •Dynamic target fallback selects best available supervision signal
- •Graph edges based on 3.0 km geodesic neighborhood radius
- •Message passing with alpha = 0.70 for 2 iterations smooths scores
Budgeted Expansion
- •Greedy facility expansion maximizes marginal demand coverage
- •Objective multiplier: 0.7 + 0.3 * learned_node_score
- •Supports fixed budget scenarios at 10, 25, and 50 sites
- •Service-radius safeguards prevent trivial full-coverage selections
Engineering Safeguards
- •Deterministic geospatial scoring for reproducible outputs
- •Graceful target fallback logic when labeled utilization is unavailable
- •Configurable hyperparameters for radius, smoothing, regularization, and budget
- •Artifactized outputs for traceability, including recommendations and clusters
- •Frontend fallback logic for sparse charger-count fields in OCM data
Current Output Snapshot
Evaluation caveat: this artifact snapshot shows coverage_ratio at 1.0 and aggregate_marginal_demand_gain at 0.0 across all scenarios. That saturation result is itself a useful technical finding and indicates the current evaluation setup needs tighter constraints or revised objective decomposition to preserve scenario discriminability.
Validation and Tradeoffs
Validation Approach
- •Schema-consistent CSV and GeoJSON artifacts consumed by UI without runtime conversions
- •Spatial checks via haversine neighborhood and radius calculations
- •Recommendations and clusters validated across scenario product paths
- •Inspectable learned weights and propagation outputs support explainability
Key Tradeoffs
- •Interpretability over black-box complexity via ridge plus graph smoothing
- •Offline deterministic optimization over expensive live recomputation
- •Coverage optimization speed versus fine-grained demand calibration
- •Hackathon velocity with clear follow-up path for production hardening
Impact and Data Disclaimer
ChargePilot demonstrates end-to-end applied optimization: multi-source geospatial ingestion, graph-based ranking, constrained expansion planning, and full-stack delivery for decision-makers. The Biggest Technical Challenge award reflects both the technical scope and execution quality under hackathon constraints.
Data disclaimer: while the system uses real public datasets such as Open Charge Map and City of Austin open data, some intermediate variables and scoring targets were synthetically generated or proxy-modeled for hackathon execution speed when ground-truth labels were unavailable. These synthetic components validated pipeline behavior and product interaction, not production-calibrated forecasting accuracy.