Harmonic Lens: AI Spectral Interpretability Engine

A mathematical "X-ray" for AI. Auditing neural bias using spectral decomposition and Gemini-powered reasoning to ensure AI safety through mathematical transparency

Comment

Inspiration

Inspiration

Here is a concise, high-impact version for your Devpost submission, specifically tailored to the "Harmonic Lens" project. Project Story Inspiration Inspired by classical optics, we aimed to solve the "Black Box" problem by treating neural layers as lenses. We replace vague heatmaps with rigorous mathematical deconstructions using Spectral Theory to see what AI truly "focuses" on. What it does Harmonic Lens decomposes neural features into Fourier, Zernike, and Chebyshev bases. It applies adaptive frequency masks to identify model blindspots, with Gemini AI providing real-time "Narrative Synth" explaining why a model’s prediction flipped. How we built it Built with React and a custom Spectral Engine. We integrated Google Gemini API (gemini-3-flash-preview) to translate energy shifts into human-readable insights and utilized Recharts for real-time frequency telemetry. Challenges we ran into Optimizing 2D-DFT and complex Zernike polynomials for in-browser execution was a hurdle. We achieved a fluid 60FPS "Reactor" UI by implementing aggressive matrix caching and optimized pixel-data pipelines. Accomplishments that we're proud of Successfully bridging raw signal processing with Generative AI. We created a "Cyber-Science" interface where rigorous math meets narrative storytelling, making deep-tech interpretability accessible to any researcher. What we learned Spectral dependency varies wildly. We discovered that many models over-rely on low-frequency global symmetries (Zernike modes), making them surprisingly vulnerable to high-frequency texture perturbations

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Harmonic Lens: AI Spectral Interpretability Engine

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Harmonic Lens: AI Spectral Interpretability Engine

Built With

  • 2d-dft
  • chebyshev
  • google-gemini-api
  • react
  • recharts
  • tailwind-css
  • typescript
  • zernike-polynomials
Share this project:

Updates