Each card covers the problem, what we built, the models and stack, and an honest status. These are previews. Prompts, agent designs and build recipes are held back for members.
Live public and in use
In daily use running in-house
Prototype built, not shipped
Research exploratory or paused
In daily use
Personal AI Operating System
A team of about 30 specialist AI agents that run research, briefings, inbox work, content and reliability checks under one set of rules.
Problem
One person running many projects needs research, monitoring, drafting and follow-up done every day, without handing an AI the keys to send, spend or publish on its own.
What we built
A control plane that routes work to specialist agents, each with a written charter, a limited set of tools, a budget ceiling and a clear stop condition. Agents share one job board and hand work to each other. Anything that leaves the system goes through a single approval queue the founder works from his phone. A dashboard shows every desk, run and failure, and an off-machine monitor raises an alarm if the system goes quiet.
Models and stack
Claude (default provider)
ChatGPT / Codex
Grok
Kimi
Cursor cloud agents
TypeScript
MCP tools
Slack and Telegram
Status
In daily use. Reliability work is ongoing: in October 2026 we traced missed morning deliveries to scheduled sign-in failures and added backstops.
In daily use
AM / PM Edge briefings
Twice-daily market and news briefings, built and checked by agents before anyone reads them.
Problem
A useful daily briefing needs fresh sources, correct numbers and an honest note when something is missing. A stale or wrong one is worse than none.
What we built
A morning and evening edition pipeline. Every figure is traced to a source, every claim gets an integrity and confidence rating, and an edition only counts as delivered when the send, the recipient-side check and a separate receipt all agree.
Models and stack
Claude
Grok
Python
Email and Telegram delivery
Status
Running in-house. Delivery checks were tightened in October 2026.
Live
The Majors Pool
A live golf majors pool site for a private group, with tournament data that refreshes on its own.
Problem
A salary-cap pool across the year's major championships needs fast, accurate scoring and real context for every pick, all tournament long.
What we built
A static web app with live leaderboards that refresh every 15 minutes during play, odds and strokes-gained context, previews, recaps, course notes and a multi-year history. Serverless functions proxy golf data and blend public sentiment signals from forums, news feeds and prediction markets. Agents work from written data-verification and presentation standards before each event.
Models and stack
Claude
Claude Code
JavaScript
GitHub Pages
Supabase Edge Functions
DataGolf API
Status
Live since March 2026 and used through the 2026 majors.
Prototype
Remi
An iOS assistant that turns plain-language notes into prioritized tasks and reminders.
Problem
To-do apps make you do the sorting. Remi takes a sentence like you'd text a friend and works out the task, the timing and what matters most.
What we built
An iPhone app with sign-in, a database secured with row-level security, storage and serverless functions. A language model parses each note into structured tasks, and a behavioral scoring engine ranks them. Built with parallel Claude Code sessions split by role: tech lead, design, backend and QA.
Models and stack
Claude Code
OpenAI gpt-4o-mini
Expo / React Native
Supabase (Postgres, Auth, Edge Functions)
SMS
Status
Version 1 is feature-complete with its test suite and security checks passing. Not yet submitted to the App Store.
An AI workspace for running a search to buy a small business.
Problem
Searching for a company to acquire means sourcing, screening, outreach and diligence across hundreds of possible targets, usually with a spreadsheet and a lot of email.
What we built
A Claude Code workspace organized as dozens of skills and agents covering the search from thesis to diligence, aimed at audiovisual, smart-home and low-voltage integrators in the Northeast. Scheduled tasks took over routine sourcing in mid-2026.
Models and stack
Claude Code
Claude Cowork scheduled tasks
Markdown knowledge base
Status
Paused since mid-2026. Targets and deal terms are private and not shown.
Prototype
INDEXX
Turns a personal library of saved social videos into a searchable, cited wiki.
Problem
A big library of saved videos is little use if you can't search what was said in them.
What we built
A local archive of saved media, cloud transcription, tags and facets, and a wiki whose answers cite the original clip. Media stays on the owner's computer; only skills and conventions are shareable.
Models and stack
Transcription models
Claude
Local archive
Wiki generation
Status
Working prototype for personal use.
In daily use
X content and analytics crew
A five-agent crew that researches, drafts and reviews posts, then reads how they performed.
Problem
Posting consistently takes ideas, hooks, a plan and an honest read of what worked. Fully automated posting is a liability.
What we built
Coordinated agents for ideas, hooks, planning, analysis and direct messages, plus a performance read on what actually landed. Nothing is posted without approval.
Models and stack
Grok
Claude
ChatGPT
Status
In use. Zero autonomous posting by design.
Research · paper only
LOOP quant desk
A research desk for trading ideas: research, code, backtest, paper trade, post-mortem.
Problem
Most trading ideas fail. The useful part is testing them honestly and remembering which ones failed and why.
What we built
One agent owns the loop from hypothesis to post-mortem, with a weekday morning brief and a ledger of failed tests. Every backtest states its data source, date range and fill assumptions. Paper trading is disarmed by default and the desk never places real-money orders.
Models and stack
Python
Claude
Grok
Market data feeds
Status
Research and paper trading only.
In daily use
Collectibles valuation desk
Values trading cards and memorabilia from completed sales, not asking prices.
Problem
Asking prices on marketplaces say little about what something is worth.
What we built
Agents that pull comparable completed sales, label where each number came from and how confident it is, and run watch routines that flag listings meeting set criteria. Models are cross-checked against each other on the same dossier.
Models and stack
Grok
ChatGPT
Completed-sale data
Status
In use for a personal collection. No buying without a person deciding.
Research
Conspiracy Ledger
A sourced, versioned ledger of conspiracy claims and what the evidence actually says.
Problem
Claims travel faster than sources. A ledger that pins each case to its evidence is more useful than a hot take.
What we built
An investigations agent with research support maintains a canonical ledger released in numbered editions, each case tied to its sources.
Models and stack
Grok
Research agents
Status
Ongoing research project.
In production
The Drought Kid
A Buffalo Bills and Sabres fan magazine produced with an agent newsroom.
Problem
A weekly fan publication needs research, sourced figures and a steady cadence, without inventing stats or quotes.
What we built
A daily research pull and a weekly issue pipeline. Every money figure carries a source and date, estimates show their method, and nothing is posted without approval. A producer agent drafts clips, images and repurposed pieces.
Models and stack
Claude
Grok
ChatGPT
Image tools
Status
Issues in production. Posting is always a human decision.
Want something like this?
We build agent workflows and custom tools for small teams.