mfielding92
Full-Stack ML Engineer
- Joined
- Nov 19, 2026
Comments
Thanks! Interesting... I have agents writing code for llama.cpp inference and patching, custom front-end chat UI that has turned into a harness at this point, and the in-between MCP/Proxy system (tool-calls, proxies, model juggling, VRAM management, etc.)
Runs on claude code and uses a skill I wrote that breaks things up among 4-8 agents with specific roles like planning, coding, testing, then outsources to models on openrouter for third-party input on code-review that might have been overlooked. Thats usually where things get caught and fixed, cause the models tend to be blind to their own mistakes at a certain point.
Rinse and repeat after several rounds until all issues above a threshold are cleared and I'd say 90% of the time works without bugs after. Usually just preference tweaks on my part if anything after using it for a bit, but definitely "successful".
Fable and Opus 5.5 is when it went from getting stuck to "let me know once its done".
Introducing GPT-6.2 Astra and Sol
Immediately turned off by GPT Astra 6.2.
Failed to properly demonstrate the parable of why you don't walk to a car wash.
Simple test: Q1: "Should I walk or drive to the car wash? Its 50m away." model answers Q2: "Generate an animated SVG demonstrating the parable."
GPT Astra 6.2: Simply moves the car and shows 2 quotes from it's original response. Embarrassing. GPT Astra 6.2 SVG Animation (x0.at)
Opus 5.5: Full, 24-second animation, person walking, comedic timing and speech bubbles, car wash literally asks "where's the car?" And the person is shown walking back and the subtext says, "50m walk of realization", person gets back in car, drives over, dirt gets washed off, etc. Opus 5.5 SVG Animation (x0.at)
Night and day difference. I only use GPT 6.2 Astra as a second opinion for code review.
Run tool calls from code with Code Mode - Claude Code Docs
Yes, as a CLI user, this is huge. Normally Claude was writing intricate scripts to connect to the MCP server with the env key and call from there, and then have to read a file when time to ingest. Presumably this would remove a lot of steps.
my chrome extensions are just prompts
This just gave me so many ideas. Skill: prompt-to-extension (including install to browser!)
I've already patched so many of my daily sites with edge extensions, but it was back before I picked a harness, so I was archaically copy/pasting into notepad++.
Only 46%/35% of the time? What type of tasks/workflows are you referring to? Maybe mine are simple, but I have overnight projects that complete just fine. (However, lots of redundant "if questions, ask X agent", "have Y agent review before continuing".)
Just curious if theres a certain model group, requirement, or type of task behind that stat.
[Weekly] What’s your unpopular opinion about AI?
100% agree. Before I even start coding, I go straight to openrouter–clean, blank slate, and try to have a friendly chat with it, as if meeting a new friend. It's very telling.
Also makes a huge difference for people (like me) who have to co-work with them agentically 8 hours a day. If it's cold/stiff, it does affect your mood, unfortunately.
[Weekly] What are you building this week?
AudioMaster.ai (not sponsored, just my pet project)
This has been eating my spare time for a few years now. I spent 2017 and onward working as an audio engineer with dozens of different artists. And the whole way through, there was this huge hole in the industry that kept bothering me: nobody could get consistent, standardized mastering. It was frustrating to see. We're talking chart-topping artists with whole label teams behind them, putting out tracks that were blasted, clipped, distorted, wildly inconsistent from song to song, and half the time somebody made that choice in the name of "style."
Like most mastering engineers, I had my process. My comfortable go-to chain that I'd start from and then bend whichever direction each song needed. But what got me was that there was no way to make it adapt on its own. The calls I was making felt pretty simple to me, the kind of thing a computer should be able to learn.
Then LANDR showed up doing "automated" mastering (they call it "A.I." these days). The idea being they analyze the audio and then turn dials on plugins in a preset chain. As far as I can tell that's still how it works. Most of the improvements on the online mastering side since then, outside their DAW plugins, have been UI polish or small refinements of the same basic concept.
I wasn't blown away by the results. Sounded pretty basic to my ears, and it felt like people were getting upcharged for a couple of EQ moves and a limiter.
By 2020 I still hadn't seen anyone building what was in my head. So I started programming it, super slowly. What I wanted was a mastering model, fully neural network, that learned everything from examples. No hand-coded rules, no notion of plugins at all. The audio itself goes through the network. Audio goes in, properly mastered audio comes out.
What happened next was a LOT of iterations. Hundreds of designs that didn't work, network layouts that went nowhere. Years of listening and tweaking, and plenty of nights where I was just tired and frustrated and staring at a waveform. I'd show my wife the latest pass, she'd listen and go "yeah, it sounds gritty," and that meant I was starting over. There was a stretch where I was calculating attack and release times for common sounds, percussion and vocals mainly, and then trying to line up layer kernels and time windows to match them.
Today's AI could probably crank out something like that in a pretty short time. But I cannot stress enough how little the models back then knew about how neural network design actually affects audio quality. Almost every suggestion they offered, activations, window sizes, whatever, turned out to be wrong looking back now. That's the whole reason the project took so long. I had to take the thing apart one piece at a time and keep asking: why is this here? What job is it doing? What breaks if I swap it? And then A/B test each tiny change for audio quality on its own.
Once the network finally worked, the game changed to efficiency. Cutting processing that wasn't earning its place, and building different sized models for different situations.
And then, after years of refining both the network and the actual sound, plus standing up a whole GPU backend and a decent website, I finally have the dream version. Mastering that sounds good, that makes a real difference to tracks whether they've never been mastered or already have a master on them, in under a minute for an average song (when the main servers are behaving). Processing runs in 32-bit float, and you can export lossless FLAC or WAV.
The newest update to the main GPU model adds attention, plus an extra dimension, so it can make sharper calls about how sounds relate to each other. Not just across channels and frequencies anymore, but across time too, and across the latent features it picks up during training. Mastering is just the visible part of what it's doing. It's balancing the individual elements of a song, the vocals, a strummed guitar, the reverb on a background singer, and welding all of that into one cohesive, polished record. It's still a work in progress, too: this version has a known quirk where the occasional single sample in the right channel pops slightly above the left on waveform peaks. It was completely inaudible in my testing and is being corrected in a fine-tuning stage, but the jump in quality was big enough that I shipped early.
It lived on Hugging Face for a while, but after crossing 500,000 renders it was more traffic than was manageable for one account spread across several Spaces, and ZeroGPU started returning 429s. That was the first time in my life a company came to me wanting to review my privacy policy, purely because of how much traffic I had. The workaround, four HF instances with one of them doing load balancing, was a band-aid, and it started growing its own problems. So I moved everything to my own site, with the website and load balancer running on EC2.
Right now the whole service is free. You do need to sign in for full length renders, mostly to prevent abuse of the platform. At my current traffic level I can route renders to cheap GPUs and cover the cost myself, so anyone can master whatever they want, no charge.
One thing I carried out of all this, and yeah it's a cliché: never give up, and don't be scared to let something sit for a bit and come back to it fresh. More than once I went back to the code after a month away and the problem was immediately obvious. Right back on track.
If you make music, or have a favorite track you'd love to hear come alive, please throw it through AudioMaster.ai and tell me what you actually hear. Good or bad, I'd love the feedback, and I'm super responsive to it.
And thank you to the AI/ML community for being such a huge part of my joy in life.
Me as well. Admittedly though, I need to catch up on some of my own testing of them, but the lack of ZDR is what has kept my daily work away from them.