Proof of work
What I've actually built.
The services on this site come out of a much larger body of work: a fleet of autonomous trading agents, a research platform with dozens of subsystems, business software, consumer apps, and the studies that came back negative. Everything here is real, most of it is running, and none of it was written by hand — I wrote the specs, made the decisions, and directed AI agents to build it.
Autonomous agents
A trading fleet that runs itself, and a manager that watches it.
Every desk below trades simulated capital inside code-enforced risk rails, journals every decision, and can only change its own strategy with written approval. One supervisor audits them all.
The Hedge Fund Manager
A supervisor agent over the whole fleet. It audits every trading desk each morning, runs research and gated backtests on weekends, files written proposals, and reports by Telegram and Notion. It has authority over analysis and none over capital — the permission file makes that physically true.
The Swarm Fund
A ten-agent maker/checker quant fund with pre-registered pass gates, a once-only holdout, and an adversarial risk officer with veto power. It tested thirteen strategy ideas and rejected all thirteen — the honest empty book that became the Autonomous Fund Kit.
Equity swing agent
A self-improving paper-trading agent: research, a model decision, code-enforced guardrails, a journal of every trade. The weekly review drafts a new strategy version that only applies after human approval, and the risk rails are off-limits to the learning loop.
Short-only swing agent
The mirror image: a regime-gated short-side desk for US equities. Its learning loop is asymmetric by design — it may tighten stops and sizing over time, but it is never allowed to loosen them.
Vol Desk
A mechanical options desk that trades off dealer gamma exposure. A rulebook executor with no AI anywhere in the runtime path, scheduled on a cloud server with Telegram reporting.
Theta Desk
A premium-selling, defined-risk options book — credit spreads and iron condors on the major index ETFs — built as a sibling of Vol Desk rather than merged into it, so each can be judged on its own.
The Wheel agent
Cash-secured puts rolling into covered calls on quality-plus-momentum names. Twenty-three modules, a 76-check self-test, a dedicated paper account, and every integration verified before it was allowed to run on a schedule.
Crypto mean-reversion agent
A mechanical failed-breakout playbook for crypto majors with fee-aware position sizing and two fee ledgers, so the breakeven spread on any venue is a documented number rather than a hope.
PatchTST forecaster
A transformer time-series model on SPY, wrapped in a six-layer risk stack so the forecast can never act alone. Built in gated phases with a review between each.
Cowboy Trader
A research bot, not a trader: it collects and credibility-scores retail trading claims from Reddit and X — more than 4,600 posts so far — to find out which playbooks survive contact with a backtest. Paper-only by design.
Edge Bot and the AI-managed fund
Two forward tests that run on a clock: a $100k simulated account trading the dashboard's swing frameworks every morning, and a $10k paper fund that a model rebalances twice a week inside a macro risk dial.
One live agent, sized at $100
The only agent trading real money, deliberately tiny. Its backtests rejected two whole engines; the single rule that survived is a 200-day trend filter on bitcoin and ether. Everything else above runs on simulated capital.
Inside the Market Analytics Dashboard
One platform, dozens of subsystems.
The free analytics platform linked under Services is the largest single thing I've built. This is what is inside it. See it live ↗
Valuation engine
A capital-allocation DCF with a forty-knob configuration panel, an eleven-method football field, Damodaran's datasets across 157 countries, Mauboussin's market-implied moat duration, FMVA stress tests, and a separate branch that values banks properly.
AI investment-committee reports
A twelve-persona committee deliberating over a deterministic dossier, rendered as a branded PDF — for stocks, crypto, and ETF holdings — sold for $5 through Stripe, plus a research template that produces PDF and PowerPoint from one spec.
Investment Desk research agent
A four-part hedge-fund-style report with a committee and a PM decision, built from filings, standing data pulls, and web research. The first heavy run produced a 28-page PDF with 24 charts.
Strategy Lab
A deterministic backtest engine behind a four-audit honesty stack (independent ledger replay, look-ahead audit, cost monotonicity) and a blind train/test split, with an AI strategist that writes specs the engine then tries to break.
Factor model suite
Seven mechanical factor portfolios rebuilt quarterly from point-in-time data, each with a full backtest record and a walk-forward paper test against the index that runs every day.
Options desk
Daily screens and a Top-5 leaderboard, a budget tab that sizes trades to account brackets, a stale-quote gate, and a preferred-stock section that extracts call terms and yield-to-call straight from EDGAR filings.
Macro suite
Fed Watch, a Recession Watch with sixteen indicators and two NBER-validated models, an economic calendar that tracks consensus drift and surprise in each indicator's own sigma, a full economy-and-market cycle study, and sector reference pages.
Research group
SEC filings and 8-Ks, AI-extracted M&A deal terms, congressional trades, and Capitol Watch — 37,326 roll calls back to 1987, floor-vote passage odds, and a map from bills to the equities they reach.
Fund star rating rebuilt
The published star-rating methodology, reconstructed for 25,726 US funds and 3.3 million fund-months back to 2004 — which showed the rating does not predict returns.
Member tooling
A Monte Carlo lab, a CFA-curriculum ratio and quality lens, one-page stock report cards, a research-assistant chat widget, site-wide paper trading, and a shared research library of more than fifty commissioned studies.
Platform plumbing
Accounts, admin console, purchase ledger and a social layer; an installable web app; a caching tier with staleness watchdogs and circuit breakers; and an X posting engine that turns live data into roughly 44 posts a week, every one of them human-approved.
Honesty as a feature
Every model-spending button is owner-locked, every number carries its basis, and features that failed their blind tests were removed from the site rather than left up to look busy.
Business software
Operating systems for real businesses.
The CFO Dashboard and Restaurant Intel platforms live under Services. These are the rest.
Sage Peak platform
A productized cash-flow and diagnostic subscription: a thirteen-week cash forecast reviewed by a person every week, QuickBooks-connected, with a human approval gate before anything reaches a client, Stripe billing, and 443 automated tests behind it.
Basecamp
A real-estate underwriting and portfolio app for the Las Vegas market, built in thirteen gated phases: a 46-table data model, 23 counter-intuitive underwriting rules implemented literally, AI memo and listing-scan agents, daily alert scans, and LP-grade PDF reporting.
Macro Brief
A standalone macro-economics subscription app with daily briefs, a month-in-review newsletter, an offline-capable installable app, and a weekend audit job that checks its own numbers.
Client-acquisition engine
A cold-outreach system with research-and-draft prompts, reply classification, call-prep briefs, a suppression list, CAN-SPAM guardrails, and a hard rule that no batch goes out without human approval.
AI-visibility audit
A working demo that measures how often AI assistants recommend a given local business, scores it against competitors, and produces a fix plan — built as a product candidate for local-business marketing.
Advisory tooling
Thirteen-week cash-flow forecast models, a proforma model suite, statement ingestion from QuickBooks PDFs, and Excel recalculation pipelines — the back office behind the finance services on this site.
AI newsroom demo
A regional Spanish-language weekly's archive — 773 articles — scraped, style-analyzed into eighteen voice cards, and rebuilt as a 1,070-page static site with an AI-newsroom demo, all in one overnight build for an owner pitch.
QRPulse
Dynamic QR codes with live scan analytics and administrator-provisioned accounts.
Consumer, family, and community apps
The range.
Not everything is finance. Some of these were built for my family, some for a hobby, some to see how fast an idea could become a working product.
Vegas MTG Singles Finder
A price aggregator for Magic: The Gathering singles across Las Vegas game stores — roughly 37,000 in-stock listings and 19,000 distinct cards, matched across stores through the Scryfall catalog.
Kickball Kids
A browser kickball game built for a young son: canvas rendering, AI fielders, base-running, unlockable rewards, and an installable app with an App Store path. Game logic is kept separate from the renderer on purpose.
Storytime
A private family story app: 140 world folktales, each written at five reading levels and illustrated in one of 37 art styles through a reviewed generation pipeline — the whole collection illustrated for about $82.
Cheapest Fare Finder
A flight-search prototype whose only job is the cheapest fare between two cities on given dates. No ads, no upsells, no sign-up wall — built on a swappable provider interface, currently against test data.
Gospel Study
A scripture study app with a cross-tradition reading library, verse-level cross-references, topic study, and talk preparation in three lengths, with metered AI features.
Life Dashboard
A two-seat household app for business cash, personal finances, fitness, reading, and a day schedule — with Apple Health import, a shared money engine, and a 67-week study curriculum generated and loaded by script.
CFA Level I study platform
Practice questions, mock exams, essays, notes, and progress tracking — original model-generated material with its limits stated plainly on the front page.
Euro Translator and Wedding Day Timeline
Two small tools from real trips and real events: a two-way voice-translation prototype for travelling, and an event-day scheduler with built-in travel cushions.
Research that said no
The results I publish anyway.
A system that only shows you wins is a system that cannot say no. These are the studies that came back negative, kept on the record because that is what makes the positive ones worth anything.
Thirteen alpha archetypes, no deployment-grade edge
An autonomous research harness with pre-registered specs, walk-forward testing, and a blind holdout consulted at most once per idea. The best lead — volatility-managed exposure — was plausibly real but not proven, so nothing shipped.
Two dashboard features removed
A regime monitor and a 'rocket' scanner both failed blind validation when re-tested properly. Both tabs came down.
A consensus strategy with no historical edge
The backtest of a seven-list author-consensus strategy found nothing. It runs only as a clean $10k forward paper test now, so the live record will settle it.
The star rating doesn't predict returns
Rebuilding the fund star rating for 25,726 funds made it possible to test it. It fails as a forecast, and the site says so.
'The next GameStop' — none of them
A ten-factor scorecard measured the actual 2020–21 setup and scored six candidates against it. The closest reached 52 against GameStop's 84.
Model books that don't clear the index
Of the published model portfolios, two fail to beat the index on a risk-adjusted basis. Their one-page briefs say exactly that.
How all of it gets built
Four rules that never move.
Paper first, structurally
Simulated-trading endpoints are hard-coded and asserted at startup. Going live is a human decision, never a config flag.
Pre-register, then test once
The spec is committed before the held-out data is touched. The statistical bar rises with every trial that fails.
Machines propose, humans approve
Agents draft strategy changes, outreach batches, and social posts. Nothing applies, sends, or posts without a written yes.
Ship with an audit trail
Every system keeps a change log of what it altered and why. Trust in an autonomous system comes from being able to read its history.
Want something like this built for your business?
Tell me what you're trying to do. I'll respond within one business day.