APEX / AlphaEdge — an options trading platform built like a production system

What is APEX?

APEX (AlphaEdge) is a self-built options analytics and semi-automated trading platform. It ingests real-time market data, classifies market conditions, selects strikes using an Expected Move framework, scores signals through a Bayesian expectancy engine, and manages positions via per-symbol state machines with a five-layer safety architecture. The frontend is a React 19/TypeScript/Vite SPA served at port 5175. FastAPI handles the backend at port 8010, SanthoshIAS routes market data at port 7700, and DuckDB stores everything.

It is v11. Each version is a response to something that went wrong — systematically diagnosed, root-caused, and fixed.

Why build it?

Most retail options platforms are built for the median trader. The median trader wants a simple P&L view, a basic scanner, and a one-click order ticket. That's fine for the median trader. It's not fine for someone who wants to model expectancy across market regimes, run cross-sectional z-score composites on volatility surface changes, or implement a Roth conversion optimizer on the same stack that manages live positions.

The existing tools also don't let you audit their logic. When a signal fires and the trade loses, a black-box platform gives you nothing to learn from. APEX gives you the full attribution stack — which gate failed, which regime misclassified, which exit rule triggered too early or too late. That feedback loop is the point.

What went wrong and what was fixed

APEX v10 ran a paper trading account. The postmortem identified two root causes of significant losses.

The naked-short bug. 24 illegal sell-to-open orders were executed on positions that weren't open, causing forced assignments. The estimated impact was approximately $21K in paper losses. Root cause: the position-state check was reading stale data from a cache that wasn't invalidated on fill. Three protection layers were implemented: a close-intent gate that validates position state before order submission, an HTTP 422 rejection at the order endpoint for any order that would create a net-short position in an account not approved for naked short selling, and a startup assertion that refuses to initialise if position state cannot be verified.

The give-back problem. 73% of stopped trades peaked profitably before reversing — approximately $65K in unrealized gains surrendered to stop-outs. Root cause: the exit architecture was a single trailing stop with no regime awareness. A position in a trending market and a position in a mean-reverting market need different exit logic. v11's exit architecture has four components: a break-even mover that locks in basis once a position reaches a threshold, a regime-change exit that closes positions when the market character classifier signals a regime shift, a theta manager that accelerates exits as expiration approaches and theta burn accelerates, and a tranche policy that allows partial exits at different profit targets.

v11 Architecture

Strike Selection V2. Replaced static percentage-based strike distances with Expected Move units derived from ATM straddle mids. Each candidate strike is scored on a composite of liquidity, IV rank, expected move coverage, and regime fit. Four regime presets — PINNED, TRANSITIONAL, AMPLIFIED, DEFENSIVE — adjust the composite weights. The portfolio risk envelope scales with equity, conviction score, IVR, and heat caps.

Volatility Expansion Scanner (VES). A cross-sectional z-score composite that replaces a flawed multiplicative formula. GEX (gamma exposure) is oriented as a signed signal — negative GEX means dealers are short gamma and moves are amplified. A K-of-N conviction gate requires signal agreement across a minimum number of inputs before routing to the strategy classifier, which outputs one of LONG_PREMIUM, SHORT_PREMIUM, GAMMA_MOMENTUM, or WATCH.

Bayesian expectancy engine. Rather than a simple win-rate and average-win calculation, the expectancy engine updates prior beliefs about strategy performance as new trades settle — adjusting for market regime, IV environment, and time-of-day. Strategies that performed well in trending markets are down-weighted in mean-reverting regimes. The engine informs position sizing through a Kelly-adjacent formula with a conservative fractional multiplier.

Per-symbol state machines. Each symbol under management runs a state machine — IDLE, SCANNING, ENTERED, MANAGING, EXITING — with explicit transitions and guards. No order can be submitted unless the state machine is in a valid state for that order type. This is the architectural response to the naked-short bug: position state is a first-class concept in the system, not an afterthought.

What it saves

Directly: APEX runs in paper mode before any live deployment. Every strategy failure costs paper dollars, not real ones. The $86K+ in diagnosed paper losses represents systematic failures that were caught, root-caused, and fixed before touching a live account. That's the purpose.

More importantly: APEX is where quantitative ideas meet production-engineering standards. The same discipline — five-layer safety, state machines, Bayesian attribution — is what makes the wire automation platform reliable at $400B+ in volume. One is personal; the other is institutional. The engineering philosophy is identical.