Work

Problem

I wanted one library that fills itself, cleans up after itself, and plays on every screen I own, including my phone away from home and the Xbox on the living room Wi-Fi. Managing files by hand was never going to happen.

Built

• Jellyfin is the front end. Radarr, Sonarr and Lidarr manage movies, TV and music, Prowlarr handles indexing and Bazarr handles subtitles. Downloads run through qBittorrent inside a VPN container. A Soulseek client with soularr upgrades lossy albums to FLAC.
• Scheduled tasks do the housekeeping: a nightly search for missing titles (Radarr and Sonarr never re-search on their own), a sync every 15 minutes so anything deleted in Jellyfin does not come back, and a daily sweep that retires lossy files once a lossless copy lands.
• Music got its own pipeline. Exports from four streaming services are matched against the library every morning to rebuild playlists, and each unique album is looked up and monitored in Lidarr so the collection fills in over time.
• Remote access runs over a private mesh VPN with HTTPS, because the phone refuses plain HTTP.

Stack

• Jellyfin
• Radarr
• Sonarr
• Lidarr
• Prowlarr
• Bazarr
• qBittorrent
• gluetun
• slskd
• soularr
• Docker Compose
• Tailscale
• PowerShell
• Python
• Task Scheduler

Problem

I wanted a model I could talk to about anything without it living on someone else's server, with web search when I need it, and with enough context that it does not fall over on a long question. A first attempt in June jammed a 32B model against the VRAM ceiling at 8K context, and the first web search overflowed it.

Built

• Ollama serves one 27B model (a Qwen build, Q4_K_M, about 17 GB) with a 32K context window. Sampling defaults and a direct-answer system prompt are baked into a Modelfile so every client gets the same behavior.
• Preset modes are separate Modelfiles that share the same weights, so they cost no extra disk: a study tutor for my security coursework, a plain-answers mode, and a few more for specific jobs.
• Open WebUI in Docker is the chat front end. A SearXNG container gives it web search, capped at three results per query so a search can never overflow the context again.
• The GPU is pinned by device ID so the model never loads onto the integrated graphics, which was the cause of a 1 to 2 tokens per second bug. The server starts windowless at logon.

Stack

• Ollama
• Qwen 27B (Q4_K_M)
• Open WebUI
• SearXNG
• Docker
• CUDA
• Tailscale
• Modelfiles
• Windows startup scripts

Problem

Screenshots of the ComfyUI interface come back blank on my machine, and stealing focus breaks whatever else is running. I needed image generation that an agent can drive end to end, judge on its own, and take all the way to paper at true print resolution.

Built

• ComfyUI runs as a local service and is driven through its HTTP API by a self-contained runner: it queues a graph, waits, downloads the PNG and prints the path. One call does the whole job, so a dropped connection cannot leave a half-finished batch.
• A quality ladder: base generation at about one megapixel, a detail pass where faces or fine lines need it, then a model upscale to 3900 by 5700 pixels, which is 13 by 19 inches at 300 DPI.
• A second track makes art with no model at all. A NumPy renderer computes pieces directly at 4680 by 6840 pixels (13 by 19 at the printer's native 360 DPI): domain-warped noise for a nebula, a particle flow field, and a Julia set with orbit traps.
• Print prep tiles master files onto sheets and drives the EcoTank from a script, including the paper tray fix that ended a run of failed prints.

Stack

• ComfyUI
• SDXL (Illustrious, Juggernaut)
• FLUX
• UltimateSDUpscale
• 4x-UltraSharp
• NumPy
• Pillow
• Python
• Epson ET-8550

Problem

A laptop OLED plus three USB panels, a night-time red overlay, Windows Night Light, brightness and the system theme all had different owners. Wakes, cable blips and a brightness utility kept undoing each other, and the USB panels' driver crash-looped when commands hit them the moment they appeared.

Built

• One PowerShell service, registered as a per-user scheduled task with no elevation, owns every display setting. Sunset and sunrise flip the red overlay, Night Light, brightness and the theme in a single tick. Ctrl+Alt+S and Ctrl+Alt+R are the manual overrides.
• Anything I set by hand stays. After a wake or a cable blip the script restores my values, never the schedule's target.
• An idle blackout after 20 minutes that the first input clears, with click-through restored before the slow DDC/CI calls so the veil never eats a click.
• A 30-second settle rule: no DDC/CI to a USB panel until the display layout has been still that long. That came from reading the driver's crash dumps with the vendor's symbols.
• A rules harness with 43 rules and a set of mutants that must all fail it, plus a guard task that rewrites any scheduled job that would pop a console window.

Stack

• PowerShell 5.1
• Task Scheduler
• DDC/CI
• Windows Night Light
• Win32 overlay window
• WDF crash dump analysis
• Twinkle Tray (coexistence)

Problem

Vendor threads and paperwork stall when they wait on me to open a computer. I wanted the inbox to hold only what needs a decision, everything else filed and findable, and an agent doing the chasing.

Built

• A daily steward: a scheduled Claude task that reads the mail, checks the open case threads, and sends one message with only what needs me. It re-schedules itself.
• Routing rules on all six mail accounts to one spec: family and one-time codes reach the inbox, records file to a searchable folder marked read, bulk mail is never seen. Text messages forward from the phone into a mail label so the same agent reads them.
• Rules for the agents themselves: a machine rules file kept under 200 lines, a dated changelog, a validator with 13 regression checks that fails on any drift, and an AI-tell scanner that anything I share has to pass.
• Machine hygiene the agents follow: nothing is deleted, it is staged with a manifest and purged after 30 days. Background jobs start through a launcher that never shows a window. Every replaced thing is retired in the same session.

Stack

• Claude scheduled tasks
• Claude Code
• MCP servers
• Mail APIs
• Outlook COM automation
• PowerShell
• Python
• Task Scheduler

Problem

I already knew systems and business. I wanted the computer science underneath the tools I was using, from data prep to models to the code that glues them together.

Built

• Data-Centric AI: feature engineering in notebooks against real datasets, including one module with a 49,628-file image set.
• Natural Language Processing: module work in notebooks, run locally and in Colab.
• Python: log-processing programs built around comprehensions, generators, function pipelines and dispatch tables, kept close to the instructor's examples on purpose.
• Also in the program: an intro to computer vision, a capstone, and security coursework alongside.

Stack

• Python
• Jupyter notebooks
• Google Colab
• Canvas