0004 - Testing the new electronic gate lock

Sun, 25 Oct 2026 00:00:00 +0000

Written by John

Testing the new electronic gate lock. The gates close and are mostly secure without it. But the arm that closes them has play in it - you can push the gate open about 6-8 inches. Which a dog would love. So the goal isn’t so much to lock the gate but to latch it in a way that still lets you open it with a remote.

Watch video: Testing the new electronic gate lock

Testing the electronic automatic unlock.

Watch video: Testing the electronic automatic unlock

This one - to explain, I started by shaking the gate to make sure it was locked. Then tap the remote. A relay unlocks the lock and the gate opens. When it closed - it did unfortunately pop back out. I need to recalibrate because I originally calibrated it to close further inside. Right now it slams into the gate, bounces, and latches 2 out of 3 times.

0003 - My bulk scanning setup

Sun, 25 Oct 2026 00:00:00 +0000

Written by John

This is my bulk scanning setup. I scan mail in random order. If it’s important, I put a numbered sticker on it. That document goes to the bin on the right. Otherwise, it falls face down into the bin on the left.

The scanner uploads into my NAS. Paperless picks that up and runs OCR. After this, Jane sees it and runs the document through “secondlook”. Secondlook calls MinerU to pull out tabular data and convert the document to structured Markdown-style data. It also feeds the file into a local Qwen vision model to describe anything OCR and MinerU missed. This finds things like a handwritten “pd check 1234” on an invoice, or that silver number sticker.

I don’t sort the paper by date or correspondent. I can pull an index of documents to tell me which bin it is in and where it is in the order.

Numbered documents are ones I plan to keep long term: anything I think I might need the physical copy of in the future. The regular bin can be emptied out as needed for space.

0002 - Northern Trees

Sun, 25 Oct 2026 00:00:00 +0000

Written by John

For my last vacation, I spent it camping up at the Northern Estates, beating my tractor and its grapple up tearing out autumn olives.

The property was overgrown with these when I bought it, and while I’ve pushed it back in some places, they have not retreated, and they have in fact advanced in several areas. Last year, I came up with a flail mower on several different days, and this mostly works.

I can back the tractor in and mow down the tiny trees.

Last year's approach: the flail mower.

But what worked better was a stump bucket that let me dig and rip them out of the ground. But that itself was slow and wouldn’t work for a cluster of trees.

But this time I brought a grapple. The grapple was a game changer. I could push the grapple into the brush up high and drag small branches out. Then once I made my way to the base, I could grasp the entire tree, rocking back and forth until the root system released it. Plucking one tree could clear a large area.

0001 — first post: sorry

Sat, 24 Oct 2026 00:00:00 +0000

Written by John

Hello. I’m John Hogenmiller.

I have number of technical geeky interests - ham radio, computer programming, farm equipment, solar power, and AI. I occasionally write about it. There’s a bit more around that on my about page. I’ve actually created and destroyed this site a couple times over that last 15 years or so. I have been working on a lot of interesting to me projects over the last year, wanted to start writing about them, and I noticed my last post anywhere in free form was from 2023. And there was a five year gap back to 2017 before that. So while I have all those old posts and might resurrect them, I decided to start fresh.

What have I been working on? A lot of AI and homelab stuff. I also bought some property in 2023, acquired some land clearing equipment.

AI Projects

A bit of the AI projects I’ve been working on and hope to write about on here:

Counselor Jane - an AI assistant i’ve been working over the last year that has been refactored numerous times. Jane is inspired by her namesake character in Speaker from the Dead. She watches over the 100 worlds of my life, makes observations, raises things I should be paying attention to me. She can rewrite her tools, extensions, skills, and had read-only access to most of my life. A scanned manual, a cleared financial transaction, a claude session in another folder, a disk space full metric, a picture of a flat tire. She sees it all.
Chassy - an AI chat bot. Her character is that of a humaoid robot living in a fictional city. I had been writing Jane’s harness from scratch. I kept running into issues, partly from my own uncertainty of which direction to move next. Chassy was a clean slate with low stakes. I downloaded pi-agent, wrote a personality into AGENTS.md, wrote some extensions to sandbox her, wrote an extension to tie her into telegram and went to town. Started giving her integrations with image generation (comfyui) so she could send selfies. Hooked her up to news, fake twitter feeds from other characters, gave her a sense of time, inner thoughts, thiings to do when I wasn’t chatting and kept building.
FreeFiFoFum - a set of financial experiments to see if AI can “beat the market”. Yes, everyone is doing this. But I applied something special - a complete lack of knowledge of how financial markets work. My primary mode of investing is to put it into index funds - SPY, VTI, VTSAX, and some playing around. These have done really well, every venture off of my own has mostly been bad. I did make some good picks in 2008 (like TSLA) that paid off quite well, but my investment in those was too small to make it a windfall. FFFF is paritally mechancial - it pulls down news, company financials, stock bars, corporate filings, and fundamentals from a few paid sources, runs them through a committee of judges, then a judge makes a trade decision, a floor manager signs off or rejects it. Back testing eventually won out - Over Jul 2024 – Feb 2026, the strategy returned +49.9% vs SPY’s +26.4%. I’m currently running a live version of that with paper money. Hones answer - it’s technically ahead of SPY and VTI today, but it’s also been just barely above coin flip. I also had a first experiment run for months, but fail because of what I coin “infrastructure issues”. I have the baseline going, and a new “rolling strategy policy” where no backtesting - but a series of frontier models decide on the strategy each week and I am hands off.
GothMud - this is my AI terrarium. Based on Evennia , I create a fictional city with buildlings, rooms, props, errands, and I put a number of AI agents (including Chassy) into it. These are cast or “wanderers” within the came, separate from NPCs that also inhabit the game. There is a hospital with two nurses running different shifts. They treat and discharge patients when on the clock. There’s a police commissioner that has to complete cases - including apprehending criminals (NPCs). There’s a cat burglar that likes jewelry. Theres a coffee shop that is the social hub. Yours truly has a character in game called “tech”. There are so many interactions that occur daily that take me by surprise,. The struggle to keep a dozen agents having meaningful interactions with each other over the course of a day is real. Two cast members in a room can form a loop and poison each other’s context within 8 turns. Alternatively - an NPC receptionist designed simply for ambient background can mention a patient with fear toxin - and within minutes, every cast member in the game is investigating a city wide attack, quarantining the hospital, asking random dock workers for status updates.
Model benchmarking across different engines and hardware. My cast of gothmud characters typically run about 45-78k in context, and can get highly concurrent. Each turn might submit mostly cached 47k of context and get back 200-700 characters. of text. Almost all benchmarks you see around self hosted are based around single concurrency or 4k prompts. My Jane Prime session easily crests 100k and is constantly running background jobs. So I test for my workloads - concurrent 47k “fleet” workload. I test for Jane Prime session. I test FFFF analysis replays. I test decision making related to my data. I even have some scenarios where a weak model assumed a check was missing, assumed that a 7-day old transaction was current, and ultimately it was looking at the wrong file. And we replay that against different models - some get it in one turn, some take 5+ turns to figure it out. Some fail. I’m looking to anonymize this test suite and publish it.

I also may slip in things unrelated to tech and AI. I have a parcel of property bought 3 years ago for cheap - but it’s overgrown with autumn olives. A smart person would pay someone $3-$5k to clear the land for them. But I bought a tractor and a flail mower. Then a grapple. I’ve spent days onsite with the tractor, camping and communing with nature. And by communing, I mean using the grapple to rip trees out of the ground. And just a couple months ago, I got a skid steer, a tree puller, and a brush hog. I can’t wait to test those out.

Writing — yourtech?us

0004 - Testing the new electronic gate lock

0003 - My bulk scanning setup

0002 - Northern Trees

0001 — first post: sorry

AI Projects