September started with a technical question that felt deceptively simple: can we run Whisper transcription on Intel integrated graphics without a discrete GPU, and make it fast enough to actually ship? The short answer was yes, but the path to that answer required a massive injection of new code and a complete rethinking of how the stack handles hardware acceleration. By the time the month ended, I had moved the project from a prototype that needs expensive hardware to a deployable reality that runs on affordable NUCs. It was a month defined by closing the gap between experimental code and usable product, even if the middle of the month got stuck in the plumbing.
The early days of September were dominated by the iGPU work in unicorn-amanuensis. I spent the first few days chasing the SYCL implementation for Intel's integrated graphics. This wasn't a small patch. The codebase absorbed the entire whisper-cpp-igpu submodule along with custom server wrappers. One commit touched 1,130 files, and another added over 10,000 lines of Dockerfiles, converters, and server code. The performance numbers are where things get interesting. The commit messages claim 80x realtime on the SYCL implementation, while a separate production server hit between 7x and 11x realtime. I am going to trust those numbers, even if 80x on an iGPU sounds almost too good to be true. Even the conservative 7x figure changes the cost equation significantly. It means we can run this stack on a $200 NUC instead of a $2,000 GPU.
The infrastructure around that core capability grew quickly. I built out the full production deployment story, including Dockerfiles for every variant: lightweight, IGPU, INT8, OpenVINO, and comprehensive. The INT8 servers are worth calling out specifically. Quantized models run at 8-bit precision instead of 16-bit, cutting memory usage roughly in half with minimal quality loss. Combined with the SYCL backend, that makes the iGPU path viable on hardware that doesn’t have a lot of VRAM. I also built a device_manager.py that ensures when you have both an iGPU and a discrete GPU, the system routes transcription intelligently rather than just throwing everything at the biggest GPU. By Tuesday, I had the Intel integrated GPU SYCL transcription working with real-time streaming on the UC1 Pro deploy. That is the hardware path people actually want to run on. It was a bigger lift than I expected, but it finally removed the NVIDIA requirement.
The early days of September were dominated by the iGPU work in unicorn-amanuensis.
While the transcription engine was stabilizing, the surrounding ecosystem had to catch up. I built out the infrastructure layer for tool servers in center-deep-pro, adding the ability to run specialized research and search tasks as dedicated services. This involved 17 files covering Docker setup, OpenWebUI integration hooks, and actual tool implementations for academic research and report generation. It was a big chunk of work, almost 3,000 lines added across configuration files and Python modules. I also pushed a massive update to uc-cloud with 74,000 lines across 283 files, mostly focused on the Authentik SSO integration. This included full setup guides, OIDC configuration docs, and all the deployment documentation for tying Center Deep into Authentik for single sign-on. Wednesday was one of those days where three completely different kinds of work happened at once. Infrastructure for a new capability, a massive documentation and SSO integration push, and a full website deployment for unicorncommander.com, which landed in the repo with 1,137 files and about 146,000 lines. Each one alone would be a solid day. Doing all three in five commits is just how it goes when you are the one who has to make it happen.
The week of September 12 was quiet until it wasn't. I only logged activity on Friday, but that single day carried the weight of the entire week. The goal was simple: untangle the SSO setup for Open-WebUI and Center Deep Pro. What started as a quick check-in turned into a deep dive into authentication, key rotation, and the messy reality of distributed systems. The keys were not rotating correctly. The configs were lying to each other. It is a common problem. You build something, it works once, and then you forget how it actually holds itself together. I needed to get it straight so the next person does not have to dig through the same graveyard of broken scripts.
I started by documenting what was actually supposed to happen, adding an 86-line SSO configuration guide to center-deep-pro. Then I hit the real wall: the JWKS endpoint. This is the part that handles the public keys for authentication. Every service had its own idea of where the keys lived or how they were formatted. I pushed a fix to uc-cloud that touched 11 files and added 1,047 lines. Most of that was scripts and config updates. I wrote a script to configure Authentik programmatically because manual setup is how you get inconsistent environments. I fixed the ops-center auth script, updated the docker-compose files for direct, prod, and traefik setups, and tweaked the nginx proxies. The work was not elegant. It was not poetry. It was just making the scripts do the work. The setup script for ops-center now handles the heavy lifting. The fix script for the JWKS endpoint ensures the keys are fresh. The dry humor of the week is that I spent most of Friday fixing a path mismatch. The kind of thing that should have been caught in the first hour. But it is not about being smart. It is about being thorough. The scripts now handle the key rotation. The nginx configs are correct. The documentation is clear. That is the win.
The final week of September was about presentation and cleanup. I spent time untangling the magicunicorn.tech portfolio, replacing placeholder assets with real screenshots from GitHub repositories. The file structure was a mess, so I renamed the assets directory to fix a 403 error on nginx. The blog was a different kind of problem. I had generated over 1.3 million lines of text across thousands of files. It was too much noise for the current phase of the site. So I removed the blog section entirely from the navigation and the build. It is not gone forever, just paused. I needed the site to feel cleaner while I figure out what is actually worth publishing. The commit deleted 5,353 files to clear the clutter. It felt good to cut the dead weight. I also implemented a cycling headline and tagline pair on the homepage and added a lightbox for portfolio screenshots. The lightbox was a small change, but it made the images feel intentional rather than bolted on. The week moved from a single file upload to a cleaner site that actually shows what I do. That is progress.
September was a month of heavy lifting disguised as incremental updates. We moved from asking if something could work on integrated graphics to shipping the full stack to do it. We added the infrastructure for serious research tools and finally got the marketing site into version control. We also spent a day untangling the SSO plumbing so the system doesn't break the next time someone touches a config file. The result is a system that is actually usable on affordable hardware, with the plumbing to support real users. It is not perfect, but it is stable. And for now, that is enough.
Real product captures — click any to enlarge.