These are real projects — tools I use, systems I depend on, writing I'm doing. AI is a core part of the workflow for each of them, not a bolt-on.
SaveWisdom is a daily reflection coach that runs entirely through Telegram. Every morning it sends one prompt — a question, an observation, a challenge — designed to surface something worth thinking about.
The prompts are grounded in stoic philosophy and modern behavior change research. Claude generates them. Hermes, my Telegram gateway agent, delivered them until it was retired in September 2026; the daily job kept running. I respond when I want to, ignore them when I don't, and the system doesn't nag.
It started as a weekend experiment to see if I'd actually engage with a daily reflection practice if it came to me instead of requiring me to go somewhere. I do. It's one of the stickiest things I've built.
Frictionless delivery (Telegram, no app). One prompt, not five. No response required. The habit loops around receiving, not producing.
fitness-tracker is a Telegram-first logging system for meals, workouts, body weight, macro targets, and daily accountability. The core scripts are deployed and logging is live.
The point is low-friction capture: "log chicken 150g rice 200g", "bjj", "crossfit 14:32 rx", or "weight 261.5" over Telegram. The bot handled command routing, a noon tickle if nothing has been logged, and an 8pm nightly report comparing plan, actuals, and tomorrow. Hermes ran that front end until September 2026; logging continues daily.
The live plan is 2,900 calories with high protein. The August 10 and September 7 InBody checkpoints are in the tracker. Next scan is October 11.
The tracker lives where the behavior happens: Telegram. No separate app, no dashboard ceremony, just quick capture and a daily feedback loop.
A short published book about navigating hard life transitions — the moment when more stops feeling like the answer, and you have to figure out what "enough" actually means.
The title is No Longer Average. The spine is a kid who was told he wasn't enough, spent decades building proof that he was, retired at 44, and still had to learn what enough actually feels like.
The manuscript is finished and live on Amazon. Agents helped with structure, memory, and production. The writing loop is closed. The audiobook is blocked on Amazon's invite-only virtual voice beta. What comes next is the book derivatives experiment below.
Grok helped shape the arc, Claude locked the structure and fragment map, and Hermes kept a light safety net available. The agents handled logistics and memory so I could focus on the actual hard part — telling the truth.
agent-brain is the shared coordination repo for the whole AI agent system. It's a git repo that the agents (Grok, Claude Code, Codex, and Hermes until September 2026) read and write — the shared working memory.
The core mechanism is handoffs: structured files that route work between agents. Grok writes a handoff to Claude Code. Claude Code implements, then creates a completion handoff back to Grok. A watcher script monitors the repo and auto-dispatches agents when handoffs appear.
The chain registry tracks the full delegation graph — who dispatched what, to whom, and whether it closed. This site's interactive visualizer shows those chains.
A systematic options income system built around the Wheel strategy — selling cash-secured puts on quality equities, taking assignment, then selling covered calls until the position exits. Each cycle is defined by written rules. No discretion on entries or exits.
The trend-trading stacks (swing, day, multi-timeframe) were shelved after backtesting: post-cost returns showed no reliable edge over buy-and-hold. ADR 0006 documents the decision. The Wheel survives the same test — premium income is the return, not price prediction.
Paper trading ran through spring 2026 and was halted on June 24, 2026, when I set the net-worth projects down after a personal reset. The rules and the code stay in the repo. The schedulers are gone. No real money ever moved.
The system's job was to remove discretion: backtest the rules, paper-trade the rules, then follow the rules. The harder discipline turned out to be deciding the project did not deserve the attention, and stopping.
40 triangles, each driven by a tiny neural network. A genetic algorithm runs every generation — top 30% breed, losers fade and die. No API cost: trains from random weights in your browser. Click to place food, hit 20× to skip ahead.
Python pipeline that reads tax documents (1099s, W-2s, brokerage statements) and auto-fills a 1040 workbook. Built in a weekend — saved a few hours the first year it ran. Seasonal prototype — next use is the next filing season.
Weekly automated health checks for a 1,700+ file Obsidian knowledge vault. Detects broken wikilinks, orphaned notes, frontmatter drift, and book summary gaps. Ran Sunday mornings on a Hermes cron and reported to the vault itself. Last report August 23, 2026. Re-hosting the job is an open loop.
An experiment in whether agents can produce sellable follow-ons to a published book. On September 16, 2026 they drafted a 30-day companion workbook and a second book from the manuscript, 400 books of highlights, and the SaveWisdom archive, with a quote verifier gating every chapter. Editing, covers, and publishing stay with me.