Meridian — AI-Powered Parking Intelligence

Inspiration

Finding parking shouldn't be the hardest part of getting to class.

As a student at The Ohio State University, I saw how frustrating campus parking could be. Students waste valuable time driving between garages, searching for available spaces, and worrying about being late to class. Even when parking availability information exists, it often provides only estimated garage-level occupancy rather than identifying individual available spaces.

That made me wonder: What if AI could see an entire parking garage, understand which spaces are occupied, and help drivers navigate directly to available parking?

That question inspired Meridian.

I wanted to create something that went beyond another parking app. My vision was to build an intelligent parking infrastructure platform that connects computer vision, real-time data, and an intuitive driver experience.

The biggest challenge was that I didn't have access to an actual parking garage, physical cameras, or expensive GPU infrastructure. Instead of allowing those limitations to stop me, I decided to build a virtual environment where I could develop, test, and demonstrate the complete system.

Meridian started as a simple idea: Eliminate the uncertainty of parking.

What it does

Meridian is an AI-powered smart parking platform that uses NVIDIA computer vision to detect vehicle occupancy and transform camera footage into actionable parking information.

Rather than requiring a physical sensor in every parking space, Meridian analyzes camera frames to determine parking-space occupancy.

The platform combines three main components:

1. AI-Powered Parking Detection

Meridian processes camera footage using an NVIDIA-based object detection pipeline. Detected vehicles are associated with individual parking spaces to estimate whether a space is occupied or whether its availability cannot be confidently determined.

The system is deliberately conservative about uncertainty. If the available visual evidence is insufficient, Meridian avoids falsely claiming that a space is open.

2. Interactive 3D Garage Simulator

We built a virtual parking garage containing 72 parking spaces monitored by eight simulated cameras.

The environment allows us to simulate vehicle movements, generate camera frames, and evaluate how our AI system responds to different parking conditions.

This provides a practical way to experiment with physical-world computer vision without installing cameras or modifying real infrastructure.

3. Real-Time Driver Application

Meridian includes a polished, map-first driver application where users can:

• Explore nearby parking garages.
• View parking occupancy information.
• Inspect individual parking spaces.
• Select a parking destination.
• Access navigation-oriented parking experiences.
• See updates as parking conditions change.

The application connects computer vision intelligence with a consumer-friendly interface, making the underlying technology useful to everyday drivers.

We also integrated NVIDIA Nemotron through Nebius Token Factory during development to explore AI-powered contextual parking recommendations.

Our long-term goal is to make finding parking as effortless as requesting a ride.

How we built it

Meridian combines AI inference, computer vision, 3D simulation, backend engineering, and frontend development into one integrated prototype.

Computer Vision and NVIDIA AI

We began with an NVIDIA SSD-based object detection baseline to identify vehicles within camera frames.

As development progressed, we explored more specialized perception approaches, including a fine-tuned Faster R-CNN ResNet-50-FPN-v2 detector trained on synthetic parking imagery.

Our perception pipeline transforms object detections into parking-space observations using camera calibration, parking-space geometry, and occupancy classification logic.

The processing architecture follows this flow:

3D Garage Simulator
        |
        v
Virtual Camera Frames
        |
        v
FastAPI Frame Ingestion
        |
        v
NVIDIA Vehicle Detection
        |
        v
Parking-Space Inference
        |
        v
Occupancy State Management
        |
        v
WebSocket Updates
        |
        v
Meridian Driver Application

An important architectural decision was to keep simulator ground-truth data separate from the actual AI inference pipeline. The perception system receives camera imagery rather than being given the simulator's answers.

3D Simulation

We used Three.js and WebGL to construct an interactive parking environment.

Our earliest prototype contained 12 spaces. We later expanded our demonstration to a larger, 72-space garage with eight virtual cameras.

The simulator generates visual observations that can be processed by our backend, while a separate evaluation system provides ground-truth information for testing.

We also developed deterministic synthetic datasets, allowing us to reproduce experiments and evaluate perception changes consistently.

Backend and Real-Time Communication

We built our backend using Python and FastAPI.

The backend handles camera-frame ingestion, inference integration, occupancy observations, API requests, and application state.

We used WebSockets to support real-time updates between the perception system and frontend.

NVIDIA Nemotron and Nebius

We successfully tested NVIDIA Nemotron-3-Nano-30B-A3B through Nebius Token Factory as part of our AI reasoning integration.

We selected the Nemotron Nano variant because its mixture-of-experts architecture offers an appealing balance between reasoning capability and inference efficiency.

This allowed us to explore how contextual AI reasoning could complement the computer vision pipeline.

Our final demonstration also supported fallback behavior rather than depending entirely on continuous live Nemotron inference.

Frontend and User Experience

We developed the driver-facing experience using modern web technologies, including React and TypeScript.

The interface was designed around a simple principle: users shouldn't need to understand the technical complexity behind AI parking detection.

We focused on a clean, mobile-first design with garage exploration, parking details, space-level observations, and an intuitive navigation experience.

Testing and Development

We used automated testing throughout development, including:

• Pytest for backend and perception validation.
• Playwright for end-to-end application testing.
• Deterministic simulation datasets.
• CPU inference benchmarks.
• Occupancy safety and uncertainty tests.

We also used Claude Code, OpenAI Codex, and ChatGPT to support development, debugging, architectural decisions, and technical research.

Challenges we ran into

Building Physical AI Without Physical Infrastructure

Our biggest challenge was developing a physical-world AI system without access to actual parking garage cameras.

We needed a way to generate realistic camera observations, simulate changing parking conditions, and evaluate our algorithms.

Building the 3D simulation environment became a project of its own, but it ultimately allowed us to demonstrate the system without requiring a physical installation.

Detecting Vehicles Isn't the Same as Detecting Parking Availability

Object detection can identify vehicles, but that doesn't automatically tell us which parking spaces are available.

We had to develop logic for associating vehicle detections with specific parking-space geometry.

Camera perspective, vehicle positioning, occlusion, partial visibility, and uncertain predictions all created additional engineering challenges.

One of our most important decisions was to distinguish between confirmed information and insufficient evidence rather than treating every undetected vehicle as an open parking space.

AI Performance and Compute Limitations

Early development relied heavily on local CPU inference.

Our initial NVIDIA SSD baseline averaged approximately 203 milliseconds per frame, highlighting the computational challenges of processing multiple camera streams.

We explored improved detector architectures and model optimization to better understand the trade-offs between inference speed and detection accuracy.

Bridging Simulation and Reality

A model that performs well on synthetic imagery isn't automatically ready for real parking garages.

Real environments introduce changing lighting, reflections, weather, shadows, unusual vehicle positions, and camera obstructions.

Understanding this simulation-to-reality gap became one of the most important parts of our development process.

Integrating Everything as a Solo Developer

Meridian required work across computer vision, 3D graphics, backend systems, frontend design, testing, and AI infrastructure.

Getting these components to communicate correctly was challenging, particularly under a hackathon deadline.

The project taught me how important modular architecture, debugging, and disciplined testing are when building complex AI systems.

Accomplishments that we're proud of

I'm especially proud that Meridian evolved from an idea into a working, connected AI prototype.

Some of our biggest accomplishments include:

Building a 72-space virtual parking garage

We developed a simulated parking environment with eight virtual cameras, allowing us to demonstrate a multi-camera parking intelligence concept without physical infrastructure.

Fine-tuning a vehicle detection model

Our Faster R-CNN model achieved the following results on a held-out synthetic evaluation dataset:

Metric	Result
Precision	96.09%
Recall	99.22%
AP50	98.99%

These results demonstrated strong performance within our controlled synthetic evaluation environment, while reinforcing the need for future validation on real-world camera footage.

Connecting the complete AI pipeline

We created an end-to-end architecture that connects virtual camera frames, AI inference, parking occupancy state, real-time communication, and a driver-facing application.

Integrating NVIDIA Nemotron through Nebius

We successfully tested a hosted Nemotron inference integration through Nebius Token Factory, expanding the project beyond computer vision into contextual AI reasoning.

Creating a polished mobile experience

Instead of presenting only model predictions or a technical dashboard, we built an interface that demonstrates how students could actually interact with Meridian.

Building it independently

Perhaps my proudest accomplishment was developing Meridian as a solo builder. I had to approach the project as a computer vision engineer, infrastructure engineer, software developer, product designer, and founder.

The experience showed me how much can be accomplished when ambitious ideas are combined with persistence, experimentation, and the right development tools.

What we learned

Meridian completely changed the way I think about building AI-powered products.

Before this project, I primarily thought about AI in terms of models, predictions, and inference. Through Meridian, I learned that the model is only one component of a much larger engineering system.

A successful AI application also requires reliable data ingestion, accurate interpretation, thoughtful uncertainty handling, efficient infrastructure, testing, and an interface that makes the results meaningful to users.

I learned how to work with object detection models, fine-tune perception systems, generate synthetic datasets, build virtual testing environments, and connect inference results to real-time applications.

Working with NVIDIA models also helped me understand the practical importance of compute efficiency, inference latency, and architecture selection.

Exploring Nemotron through Nebius introduced me to hosted AI inference and how cloud-based reasoning models can complement traditional computer vision.

One of my biggest lessons was that impressive benchmark results don't automatically translate into production reliability. Even with high synthetic detection accuracy, real-world validation remains essential.

I also learned the importance of acknowledging uncertainty. For a parking application, incorrectly reporting that a space is open could waste a driver's time or create a poor user experience. Reliable AI should communicate what it knows and what it cannot confidently determine.

Finally, this project strengthened my confidence in independently researching unfamiliar technologies and turning complex ideas into working systems.

What's next for Meridian

The hackathon was only the beginning.

Our next major goal is to take Meridian beyond simulation and test it inside a real parking garage.

Phase 1: Physical Camera Integration

We plan to connect real IP cameras to the existing perception pipeline and evaluate how the system performs with actual garage footage.

This will involve camera calibration, networking, image processing, and real-world occupancy testing.

Phase 2: NVIDIA Edge AI

We want to explore deploying inference on NVIDIA Jetson devices to process camera footage closer to its source.

Edge inference could reduce bandwidth requirements and create a more scalable architecture for garages with multiple cameras.

Phase 3: University Pilot

Our first target is a potential pilot at The Ohio State University.

We hope to work with university parking stakeholders to evaluate Meridian in a limited garage environment, measure detection reliability, and understand installation and operational requirements.

Phase 4: Improve AI Reliability

We plan to expand our training and evaluation datasets with real garage footage, improve detection under occlusion and difficult lighting conditions, and develop stronger temporal tracking.

We also want to advance our AI-assisted parking recommendations by combining real-time occupancy observations with contextual reasoning.

Phase 5: Expand Beyond Universities

While Meridian began with student parking, the underlying technology has broader applications.

We envision deploying Meridian across universities, airports, hospitals, shopping centers, and other environments where parking availability is difficult to predict.

Eventually, Meridian could become part of a larger intelligent transportation infrastructure, helping organizations understand parking demand while providing drivers with a more seamless arrival experience.

Our Vision

We believe the future of parking should be intelligent, connected, and effortless.

Meridian is our first step toward making that future possible.

Meridian — Arrive Ready.

Built With

• computer-vision
• docker
• fastapi
• github
• javascript
• nvidia-ssd
• playwright
• pytest
• python
• react
• three.js
• typescript
• webgl
• websockets