Running heavy AI in production

VibeCheck

Client · SaaS

Running heavy, demanding AI audio models online and profitably: that's the challenge VibeCheck's founder brought me. A SaaS designed and built end to end: the app compares two tracks and computes their similarity score, perfect for checking a cover.

SaaSIA / MLStripe

01The problem

The real trap: these models can blow up response times and the server bill. It had to stay fast, secure and economically viable.

02What I shipped

I built a full self-service SaaS: Next.js front-end, a FastAPI back-end running TensorFlow, Stripe subscription, encrypted files, automatic cleanup, all deployed on a VPS with NGINX.

03Under the hood

The score doesn't come out of a black box: here is the pipeline the API runs on every analysis.

Spectrogram

Windowing + short-time Fourier transform

X (m, ω) = \sum_{n} x [n] w [n - m] e^{- j ω n}

Source separation

Learned spectral masking (TensorFlow)

M_{i} (t, f) = \frac{{| S_{i} |}^{2}}{\sum_{j} {| S_{j} |}^{2}}, {\hat{S}}_{i} = M_{i} ⊙ X

Comparison

Cosine similarity between embeddings

\cos θ = \frac{⟨ u, v ⟩}{‖ u ‖ \cdot ‖ v ‖}

Final score

Weighted sum (melody, harmony, timbre, rhythm)

S = 100 \cdot \sum_{k = 1}^{K} w_{k} \cdot \cos θ_{k}

04Key features

Track comparison & similarity score
Instrument & vocal separation
Analysis history
Stripe subscription · encrypted files

Result : ≈ 15 s to separate a track · 1,200+ files processed · recurring Stripe subscriptions. The more technical it gets, the more fun I have.

01 / 03

I digitalized an entire business

DreamLocation