PerfectShot AI

Inspiration

It was our first time in San Francisco. The sun was shining, the city looked beautiful, and we just wanted to capture the perfect moment. But there was one problem: both of us are terrible photographers. Every picture was either overexposed, tilted, or poorly framed. Instead of giving up, we decided to build a solution that would help anyone capture professional-quality photos with the help of AI. That is how PerfectShot AI was born.

What It Does

PerfectShot AI is an intelligent mobile photography assistant powered by deep learning. It combines computer vision, aesthetic scoring, and natural-language feedback to help users take better photos in real time.

The app uses multiple machine learning models together:

A YOLOv8 object detection model (TensorFlow Lite) identifies people, faces, and key objects in the frame.

A custom aesthetic evaluation model trained on the AVA dataset generates an aesthetic score for every frame. The score reflects composition balance, lighting, color harmony, and visual appeal.

Claude (Anthropic) through Groq AI interprets these scores and scene data to generate intelligent tips in natural language, such as “try lowering the camera for better perspective” or “move the subject toward the left third.”

Device sensors provide additional data such as orientation, roll, and ambient light to detect if the camera is level or if flash should be automatically disabled.

In real time, PerfectShot AI overlays visual guides such as the Rule of Thirds, Golden Ratio, and Center Focus on the camera viewfinder, and provides actionable AI feedback on lighting, framing, and subject placement.

The result is a camera that thinks like a professional photographer and guides you to your best shot.

How We Built It

We built an Android app in Kotlin using CameraX for live camera preview and frame analysis.

Core ML Pipeline

TensorFlow Lite for on-device inference of both YOLOv8 and our fine-tuned aesthetic model.

Custom AVA aesthetic model, trained locally using transfer learning on convolutional neural networks and exported to TFLite for mobile deployment.

Groq hardware acceleration for ultra-fast inference, allowing near real-time frame evaluation.

Claude API (Anthropic) for text-based interpretation of aesthetic data, turning raw numeric outputs into natural suggestions.

Sensor fusion using the accelerometer, gyroscope, and ambient light sensors to calculate tilt, roll, and brightness for exposure guidance.

AR overlays rendered on the preview screen to visualize composition rules directly in the camera interface.

All feedback is generated locally or through Groq-accelerated inference, ensuring privacy and instant performance.

Challenges We Ran Into

Running multiple ML models per frame while keeping the preview smooth.

Balancing subject detection, aesthetic evaluation, and natural-language reasoning in one feedback loop.

Training an aesthetic model that generalizes across lighting and composition conditions.

Aligning detected subjects with the camera’s live coordinate system and rotation.

Accomplishments We’re Proud Of

We built a fully functional AI camera assistant that evaluates aesthetics in real time.

The system uses deep learning for aesthetics, language models for feedback, and sensor fusion for precision.

The app genuinely improves image composition and helps non-photographers take better pictures instantly.

It successfully merges on-device inference and AI reasoning into a seamless user experience.

What We Learned

How to train convolutional networks on subjective datasets like AVA.

How to convert deep learning models to TensorFlow Lite for efficient mobile inference.

How to integrate Groq acceleration for low-latency ML performance.

How to couple a language model (Claude) with aesthetic scoring for interpretable guidance.

How to unify sensor data and computer vision models for contextual, intelligent camera feedback.

What’s Next for PerfectShot AI

Add voice-based real-time feedback for hands-free photography.

Build separate models for portrait, landscape, and night modes.

Add weather and sunlight prediction to recommend the best time for outdoor shots.

Launch PerfectShot AI publicly and continue fine-tuning the aesthetic model with larger and more diverse datasets.

Explore integrating the system into default camera apps with smartphone manufacturers.

Our long-term goal is simple: help everyone capture their best moments through the perfect balance of AI and creativity.

Built With

Kotlin • TensorFlow Lite • YOLOv8 • Groq AI • Claude (Anthropic) • CameraX • AVA Dataset • Android Sensors

Built With

• anthropic
• ava
• groq
• kotlin
• machine-learning
• python
• yolo