archive
Everything I make
This is where I keep the thinking I've done in public — research teardowns and the things I'm building. Most of it started as me trying to understand a system well enough to act on it, then deciding the reasoning was worth keeping. Browse it all, or filter down to what you came for.
Metsu: Personal Intelligence
A personal intelligence system that keeps AI aligned with a user's current reality across tools, decisions, and changing work.
Read ↦Catalyst Radar: Real-Time Market-Event Detection Engine
The whole Catalyst Radar project in one place: a Python engine that scans 163 perpetual markets every five minutes, ranks the unusual movers, classifies each catalyst with Claude, and fires a deterministic alert. Then I backtested 482 of its live alerts and the honest answer came back: the edge dies after the first hour. So I don't trade it by hand. The engine ships, the research ships, and the backtest that killed the trade thesis ships too. The thinking is the artifact.
Read ↦Palantir Foundry: Incident Triage Agent
Built this to impress Palantir before an interview, straight up. One session inside Foundry, their actual platform, and I shipped a full incident triage agent end to end: live data, an ontology, an AIP agent that reasons over runbooks, and a governed action behind a human approval gate. The fun part was just playing in their software, watching the data pipelines line up and the object state flip in real time while I watched. Same shape as the agent I shipped at Walmart, this time on their governance primitives.
Read ↦FDD Update Engine: AI + Deterministic Drafting for Franchise Disclosure Documents
Found a document nobody automates: the franchise disclosure document every franchisor has to refile every year, by hand, by lawyers. I built a demo to test it, and building it changed the whole idea. Attorneys don't draft 23 items from a blank page. They put last year's filing next to this year's and fix what moved. So that's the product now: upload the old FDD, drop in the new inputs, get a tracked-changes diff a lawyer can finalize. The actual workflow, not a text generator.
Read ↦Fixxr Competitive Teardown: AI Car Repair in a Startup Graveyard
A venture-grade competitive teardown of Fixxr, an AI car-repair startup I was weighing joining. The due diligence I'd want before betting my time on a team: where the category's graveyard is, what the real differentiated wedge is, and whether the traction holds up.
Read ↦Reverse Samwer Screen: 16 Foreign-Proven App Opportunities With US Whitespace
Here's the cheat code: find the apps already winning in other countries that nobody has built for the US yet. That's structural arbitrage, not guessing. I screened for sixteen foreign-proven opportunities sitting on open US whitespace, then broke down what actually makes each one transplantable.
Read ↦Catalyst Radar: A Quant Backtest of 482 Live Breakout Alerts
An adversarial study of 482 live alerts the engine actually fired. I graded all of them: conditional win rates, feature interactions, regime analysis, exit timing, plus an $8 Grok experiment to test whether the LLM layer predicts anything. The honest result: the only robust edge is a two-variable filter, the biggest lever is an exit rule and not an entry, and the LLM's direction call adds nothing measurable.
Read ↦Venture Assessment: AI for Regulated-Document Compliance
Before betting on a wedge, I wanted to know which one was actually better. So I put two against each other. FDD, the legal side, where every franchisor refiles a 23-item disclosure by hand every single year. Versus defense compliance documents, where my clearance gives me an angle. This is that assessment: market, tech, and regulatory risk on both, and why I led with FDD.
Read ↦OVERPOWER: A Browser Open-World Sandbox
A browser open-world superhero sandbox: procedural city, switchable powers, crowd NPCs, buildings you can level. No external assets, all of it generated in code. And it was one-shotted, a single prompt to Claude's Fable model. The experiment was simple: how far can one prompt take a whole open-world game? This far.
View on GitHub ↗Feasibility: An AI Agent for Remote Fleet Control Without a Unified API
Early thinking on future robotics and an AI fleet controller: could an autonomous agent run a heterogeneous hardware fleet with no unified API? The feasibility study that seeded the idea of an autonomous remote AI agent.
Read ↦Unsolved Problems for an Autonomous Remote AI Agent: Consumer Devices & Automotive
Where could an autonomous AI agent actually take over your devices and your car? I went looking for the real gaps, the control locked behind apps and proprietary billing that nobody has cracked. Then I sorted them: which are real, which are unsolved, and which are actually worth building.
Read ↦Catalyst Detection Engine v2: Peer Review & Architecture Spec
An adversarial peer review of my own detection architecture, measured directly against v1: what the first version left on the table, the assumptions the empirical record contradicts, and the v2 spec that fixes them.
Read ↦Catalyst Radar v1: Architecture Optimization Brief
My first hard pass at fixing the Catalyst Radar architecture. The v1 spec had three structural flaws: uniform polling, one LLM call per ticker, and a ranking that leaned on a lagging number. So I redesigned the loop, the prefilter, and the scoring to fix signal-to-noise and keep the cost near zero.
Read ↦MCP-to-Open-RMF: The Robotics Infrastructure Gap
The original research on the robotics-fleet problem: bridging MCP to Open-RMF so an AI agent could orchestrate a heterogeneous robot fleet. An honest read on the whitespace — and why the timing matters more than the opportunity.
Read ↦Every Data Vector for Predicting Music Virality
Investigating whether music virality has an objective, mathematical basis — cataloguing every data vector that might predict whether a track breaks out, and how much real signal each one carries.
Read ↦Multi-Timeframe Execution Engine: A Complete Architecture
Early architecture research, before I built Catalyst Radar. A single perp account can't hold opposing positions at once, so every multi-timeframe system has to solve the net-position problem first. This is the spec where I worked that out: three layers, 15M, 1H, 4H, on BTC-PERP, reconciled into one net position. The thinking that fed the build.
Read ↦PropJigs: API-First Wholesale Real-Estate Marketplace
My first real build. An API-first wholesale real-estate marketplace: it turns off-market contracts into AI deal briefs, hides the full address until a buyer makes a verified commit, and tracks intent and reputation in a graph. First time I took a niche all the way to a shipped product.
View on GitHub ↗Nothing in this category yet.