10 songs. 2 pipelines. 1 week of melody extraction.
The biggest discovery? basic-pitch transcribes the entire mix, not the melody.
Spike done. Lesson learned. Time to build. 🎶
@X@NoDAW_Studio#devcenterhacktobersprint
Finished my initial spike test, where I got audio source separation working. This shows that the core concept of my project is viable and doable Curious about the project? Ask in the comment section 👇
#BuildInPublic#hackathon@X
@DmitryBBLV@X On my laptop CPU, Demucs takes about 2.6× the song duration to separate (e.g. 14 min takes ~36 min), so I’ll need to make some serious decisions around the architecture and design to get past this from the user-experience side.
@X
@DmitryBBLV I'm glad you are able to forsee this. You may have save your database from endless requests, brute-force attacks, and partially of automated bot activity
The problem is not really about crashing or not, because everyone will eventually crash down the road. When you deeply understand the fundamentals, you aren't just helping yourself debug. You're giving the AI the exact context it needs to 10x.
AI can 10x your development speed, no doubt, but only if you actually know how to use it. Using AI tools with shaky coding fundamentals is like putting a Nitro on a tricycle. You'll move faster, but you're probably going to crash somewhere
👇
If skipped, small changes stop being small.
A simple update turns into touching multiple unrelated parts of the codebase because the system was never properly thought through from the start.
Slow down at the start. The rest moves faster because of it.
I learned this while building the frontend for a CEX platform.
I spent time mapping flows, locking down the design system, and analyzing edge cases before they became bugs.
It felt like no progress. It was the most important work I did on that project.
What if the problem with AI proctoring isn't the tech but the question it starts with?
Most systems ask: "How do we catch cheaters?"
What if we instead asked: "How do we design assessments where real understanding can't hide?"
Most AI proctoring systems treat students like suspects.
A trust-first system treats them like participants responsible for proving their own understanding.
@yabsssai Exactly, that tension between art and science often sinks projects. Science is vital in Phase 1 as architectural decisions compound; a beautiful interface on a shaky foundation fails later during scaling, and that's the hidden cost of letting "art" lead too early.