.png&w=256&q=75)
🤝 Top Collaborators
🤓 Latest Submissions
Convexity — the agent that cannot buy a wide book
Convexity is an autonomous options agent running on a $100,000 Alpaca paper account with zero live keys. Six stages, tape to exit: live WebSocket prints and OPRA quotes define the universe; a deterministic signal engine fires on volume breakouts; a screener ranks nearest-expiry option flow; a mandatory RiskGate evaluates a fresh bid/ask at order time; Alpaca's Trading API places limit buys at the ask, cancelled if unfilled in 20 seconds; and a pure-code exit stack runs profit locks, bid stops, IV-crush exits and a 15:50 ET session flatten. RiskGate is six deterministic rules on every entry path — signal, agent, or human. Hard spread caps (20% OPRA, 15% indicative, 12% on 0DTE). Priced at the ask, so worst-case notional is computed on what you'd actually pay. One working order per symbol. A daily loss lock counting realized plus unrealized, marked to the bid. Per-symbol cooldowns. A chase guard that cancels into widening markets instead of following them. If a gate fails, the trade does not get a smaller size. It gets a journal entry. On top sits a 15-second observe-analyze-act loop with five guard skills: drawdown, liquidity, churn, exposure, consistency. It can pause entries, lower concurrency, or flatten. It can never arm, trade, or raise a limit. It runs with or without the LLM, so a hung model can never strand a position. What it got wrong is the most useful thing it produced. On September 2 the book lost money on a day when every single trade was legal. All fifteen trade-level gates passed, because each one checks a trade. Nothing checked the portfolio. Every open position was a short call spread — three separate bets that nothing rallies — and the market rallied. Gates that check trades cannot see a book. That is the difference between a rule and a risk system. The desk is live and read-only at convexity.faruqueamin.com/desk. Paper only.
4 Sep 2026
Verdikt Know Your Claim's Odds Before You File It
When someone files an insurance claim, a car accident, storm damage, a canceled trip, a hospital bill, they're usually doing it for the first time, under stress, with no idea what "good enough" documentation looks like. They find out their claim was weak only after it's rejected or lowballed, weeks later, when it's often too late to fix. Nearly every AI tool in this space is built for the insurer's side of the table, fraud detection, claims routing, adjuster productivity. Nothing meaningfully sits on the claimant's side and says, before they hit submit: here's what's missing, here's what will get you rejected, here's what to fix first. The Solution Verdikt is a claimant-side, insurer-agnostic pre-submission audit tool. A user selects their claim type (Auto, Home/Property, or Health), describes their incident, and uploads whatever evidence they have. The app: Analyzes each uploaded document/photo against a checklist of what that claim type typically requires Produces a plain-language readiness score (0–100) with a ranked list of specific gaps Drafts a clear, factual incident narrative from the user's own description Bundles everything into a downloadable submission-ready packet Verdikt never files the claim or contacts the insurer, it exists purely to make sure the person's own submission is as strong as it can be before it goes anywhere. Every screen carries a clear disclaimer: this is general guidance based on common industry practice, not legal or insurance advice, and requirements vary by state, country, and policy.
10 Aug 2026
