Finished code with a terminal for frontend and one for backend running simultaneously
Me with my Hello Kitty Blanket

EchoElephant Denoiser — Hackathon Journey

Inspiration

I came into this hackathon as a first time hacker and freshman, honestly not expecting much out of myself and as anxious as can be. I didn’t have a team and I had multiple occasions where I thought I did but then something would happen like someone would leave, or join another team, or the team would be full. At one point I seriously considered dropping out because it felt like I was already behind everyone else and I didn't think I'd make it on my own.

But I stayed.

I decided why not just see how far I can get. And I'm glad I did. I've learned so much through this entire event. Not only about coding, but about myself. I learned that no matter how scary or difficult a situation may seem, there's no harm in trying. Even if I fail, I should still be proud I tried. I also learned a lot about coding. This was, believe it or not, my first time working frontend on a project. And man did I enjoy it. Having the freedom and creativity to make my app look however I wanted was so exciting and freeing. Whether I win something or not, I am still proud of myself.

The elephant challenge was very interesting to me. I'm someone who absolutely loves animals and wildlife, so getting the opportunity to build something that could potentially help them while also building my programming skills seemed like a chance I couldn't just pass up.

So, I created EchoElephant, an elephant call isolator. The goal of this project was to isolate elephant vocalizations from real-world noisy environments, that way researchers can analyze the audios more effectively.

What I Built

I built a full-stack machine learning web application that:

Takes an uploaded audio file
Converts it into a spectrogram representation
Uses a trained U-Net model to predict a noise-reduction mask
Applies signal processing techniques to isolate meaningful audio features
Reconstructs a cleaned audio file for playback

Mathematically, the core idea can be expressed as:

[ \hat{M}(f, t) = f_\theta(X(f, t)) ]

where:

(X(f, t)) is the noisy spectrogram
(f_\theta) is the neural network (U-Net)
(\hat{M}(f, t)) is the predicted clean mask

The final cleaned spectrogram is computed as:

[ X_{\text{clean}}(f, t) = X(f, t) \cdot \hat{M}(f, t) ]

How I Built It

Backend (FastAPI + PyTorch)

I built a Python backend using FastAPI, which handles:

File uploads
Audio preprocessing (librosa, NumPy)
Model inference using a PyTorch U-Net
Reconstruction of the final audio output

Audio Pipeline

Load audio file
Convert waveform → spectrogram
Apply log scaling for stability
Run inference on the spectrogram
Apply post-processing (smoothing, clipping, filtering)
Convert back to waveform

Frontend (React)

The frontend is a simple but functional React app:

Upload audio files
Trigger backend processing
Display and play cleaned audio output

Challenges I Faced

This project pushed me more than I expected.

1. Doing Everything Alone

Not having a team meant I had to:

Design the ML pipeline
Build the backend API
Build the frontend UI
Debug deployment and feature issues -Create the presentation, demo video, and elevator pitche all on my own

There were moments where everything broke at once and I had no one to ask. But I kept pushing through until I got a project I was proud of.

2. Being a First Time Hacker

Not having any prior experience made it that much more difficult:

I was lost the beginning, not knowing where everything was or even how to check in
Feeling anxious because of all the work I had to do alone in a short period of time
Having less experience than everyone around me
Not feeling like I had someone to talk to and ask questions

3. Audio Processing Instability In Project

The model pipeline was sensitive:

Spectrogram shape mismatches caused crashes
Some runs produced empty audio outputs
The system occasionally hung during inference I had to debug step-by-step through:
tensor shapes
masking logic
reconstruction pipeline

4. Environment Issues

Even running the backend was a challenge:

missing dependencies (torch, soundfile)
virtual environment confusion
uvicorn import errors

5. Adding More Made It Worse

As I was getting deeper into the project, adding stuff just broke other things:

I have a good prototype down so I want to add another feature then boom, nothing works anymore
Having to go back a debug the same code over and over
Overcomplicating things forced me to go to simpler models

I spent a surprising amount of time just getting the system to run consistently.

What I Learned

This project taught me more than any class so far:

How full-stack ML systems actually work in practice
How fragile real-world pipelines can be
How many different and fun things you can do with the UI
How to debug across frontend, backend, and model layers simultaneously
How much of a lifesaver github is
It's better to be proud of yourself for failing and trying then never trying at all
Most importantly: how to keep going even when you're alone and everything breaks

Final Reflection

I didn’t finish this project as perfectly as I imagined when I started.

But I finished it all by myself.

And somewhere between debugging Python errors at 2 AM and fixing React hook crashes, I learned that building real systems isn’t about everything working smoothly, it’s about staying in the fight long enough to make something real.

Even if it’s imperfect.

Even if you’re alone.

Even if it almost doesn’t work.