Slerp.audio – VDJ with WebGL2 and real-time DSP

I wanted to push the limits of what I could build with Cursor. I did not write a single line of code, only prompts, plan docs, and agent files, while directing the LLM to use automated code checks (file length, linting, unit tests, performance tests, other safe guards around gh + aws cli)

I built a browser VDJ-style demo with fullscreen WebGL2 shaders. It’s something I’ve always wanted to build but never had the time to learn all that was required. Cursor almost feels like the early days of my career, where apps felt like they had unlimited potential and I just wanted to build things. I have never done DSP programming before.

I used Cursor to build the audio pipelines that send data to the shaders and to run FFT in an AudioWorklet. Tone shaping uses native Web Audio biquads: high-pass → five-band EQ → low-pass, then into the limiter/safety chain.

I created specific workflows to allow Cursor to iterate on UI/UX (using Playwright). I also used workflows where Cursor would iterate on performance. I tuned the rendering and audiolet performance directly in Chrome DevTools, passing the profiles directly back into Cursor

Below are are some stats on the project.

Project Timeline: 13 days

Cursor requests: 1,344 Total: 814 on-demand (paid) + 528 subscription-included + 2 errored (no charge)

on-demand spend: $1,146.21

total tokens: 4.37 B (4.20 B input cache reads; 83 M input w/o cache; 62 M w/ cache write; 18.8 M out) input token cache: 96.7% of input were cache reads — long threads, not chopped chats

tree size: ~39.3k LOC (web/ TS+CSS, IaC, audio worklet, various experiments)

long threads: 537 user prompts and 7,025 assistant messages across 9 top-level agent threads