​

How I used small local models to generate coding briefs, semantically highlight inputs, and enrich NotFinalCode without letting AI become the source of truth.

I started from a question about small local decision models and data ingestion, then ended up benchmarking Laya against Kev-4B on CPU. Kev was far more accurate on my DBpedia 14 fixture, while Laya was dramatically faster under my current harness—raising a more interesting question than “which model wins?”: how much intelligence do we actually need to spend on each decision?

I tried Arch for the swagg, chose Fedora for the stability, and then spent two days customizing nearly everything anyway.

I benchmarked LM Studio and Unsloth Studio inside a VM on my Minisforum UM890 Pro. The same Qwen3.5 model went from 13.01 to 22.24 tok/s, while Unsloth’s dynamic Q4 variant reached a 43.06 tok/s median—with enough variation to make that number slightly suspicious.

How I learned QGIS through real telecom problems at Orange, trained field teams with QField, and turned recurring geographic analyses into GeoPandas, Streamlit, and Folium applications for non-technical users.

A sarcastic but honest reflection on academic success, job-market rejection, impostor syndrome, and how failing to follow the expected path pushed me to launch Uncreative Projects and start building with data, AI, OpenRouter, and Paperclip AI.

I bought a Minisforum UM890 Pro to run local LLMs. It ended up hosting Proxmox, AI workflows, search tools, agents and three game servers—while still lacking the GPU power needed for the models I actually wanted.

From learning Java with BlueJ at 18 and reinstalling Linux seven times to building BI and GIS tools at Orange, running a homelab and experimenting with AI: this is the story of how I learned that being a developer is not about already knowing the answer.