Topics

• Data engineering 29 posts

  Microsoft Fabric, Azure data platforms, lakehouse design, ETL migrations, data modeling, incremental loads, watermarks, auditability, and the boring control layers that make pipelines reliable.

• AI-native systems 15 posts

  How AI fits into real workflows without turning everything into a black box — especially systems where AI assists, but deterministic logic still owns the final action.

• Local AI & builder infrastructure 4 posts

  Ollama, OpenWebUI, SearXNG, local models, homelabs, GPUs, self-hosted tools, and the tradeoffs between local control and cloud convenience.

• Systems tinkering 13 posts

  3D printing, Raspberry Pi, SBCs, OpenWrt, pfSense, Proxmox, Unraid, embedded devices, LTE/5G modules, routers, sensors, and other useful rabbit holes.

• Product architecture 11 posts

  Small tools, solo SaaS ideas, anti-bloat thinking, workflow design, and the line where a useful feature starts turning into a platform.

• Knowledge systems 6 posts

  Markdown, search, documents, structured notes, versioning, retrieval, and the problem of keeping information useful after the initial excitement wears off.

• Applied experiments 18 posts

  Domain-specific systems: recipe versioning, genomics data, smart agriculture, multilingual media, privacy tooling, and other workflows that are really software problems in disguise.

Series

A few throughlines worth reading in order — the same idea chased across different projects.

• Data failures that stayed green

  Pipelines that reported success while producing wrong data — and the cheap, boring checks that would have caught them.

  1. The Missing Column That Didn't Crash
  2. The Keys That Never Matched
  3. Checksum, Not Filename
  4. Timestamps Beat a Mutable Status Column
  5. The Six Months I Had to Reload

• Hardware the spec sheet lied about

  What the datasheet number hides, learned by buying the wrong thing first.

  1. The NVMe Specs That Actually Matter for a Homelab
  2. I Built Cellular Routers Every Way. Then I Bought One.
  3. Everything With a Battery Cooks in a Parked Car
  4. Not Every FLAC Is Lossless

• AI, and the work you can check

  Where automation is safe, where it silently lies, and why the verification boundary is the whole design.

  1. AI Is Good at the Work You Can Check
  2. Four Results That Looked Better Than They Were
  3. Three Graphics Bugs in an AI-Written WebGL App
  4. You Can Check a Cookie
  5. Why Evidence Weight and Agreement Need Separate Scores

• Day two, when the demo has to run

  The operational maturity that decides whether a system survives past the launch it looked great at.

  1. The Happy Path Ends at Day Two
  2. Cron Is a Clock, Not a Scheduler
  3. Fail Closed, Not Fast
  4. How Twelve Rejected Files Became Unreloadable
  5. Keep the Runner Dumb

• Where the GPU stops paying

  A port that was a wash, then nine measured workloads that show where transfers, fallbacks, and index builds decide the answer.

  1. The GPU Collision Detector That Didn't Get Faster
  2. The Projects Where Data Movement Cost More Than Compute
  3. When Does GPU Parquet Actually Pay Off?
  4. DuckDB, Polars, and cuDF on One Analytical Pipeline
  5. How Much of a Spark Plan Actually Runs on the GPU?
  6. How Many Queries Pay for a Vector Index?
  7. 100 Million Points on One GPU
  8. I Moved an Entire ML Pipeline to the GPU
  9. I Wrote the Same GPU Operation Six Ways
  10. The Radio Won't Wait for Your FFT
  11. Five 4K60 Streams Were Fine. The Sixth Was Not.

• Loading files you do not control

  From incremental-load patterns to a preflight check and a full intake path: what breaks when someone else owns the file.

  1. Incremental Load Is Not One Thing
  2. The Files That Break Your Bulk Load
  3. The Limits I Built Into willitload
  4. I Tried to Make Structured File Intake Boring

• Real data on small machines

  Fabric without Spark, DuckDB on one node, and a browser as a data workbench: how far a single process goes before a cluster earns its keep.

  1. The Cluster Was Idle the Whole Time
  2. Fabric Without Spark
  3. Deploying Fabric Without a Debugger
  4. DuckLake on One Machine
  5. The Data Was Already in the Browser

• Memory and evidence around a model

  Designs that keep the model stateless and put storage, citations, and review in code the reader can inspect.

  1. Building Durable Memory Around a Stateless Model
  2. What Changed When Models Could Answer Each Other
  3. Every Answer Cites Its Source
  4. A Web Interface for the NHTSA Evidence System
  5. I Built Automatic Backlinks Before Testing the Writing

• Proxmox storage, in the order you meet it

  Choosing a filesystem, building the pool, naming the disks, passing them through, and eventually replacing them.

  1. ZFS, BTRFS, and LVM in the Homelab: Picking a Filesystem Without Regret
  2. Creating a ZFS storage pool in Proxmox
  3. Identifying ZFS disks with /dev/disk/by-id/
  4. LVM Thin Pools in Proxmox
  5. Passing physical disks to Proxmox VMs
  6. Migrating a ZFS pool to new drives in Proxmox

• Self-hosting local AI

  Why I run it at home, what the first box taught me, and what a 72 GB mini-PC changed.

  1. Why I Self-Host
  2. Building a Local LLM Box That Doesn't Need Babysitting
  3. 72 GB of VRAM, No Graphics Card

Specific tags