Trading operations / AI planning / broker-safe execution

Glyph trading operations platform

Glyph is a broker-connected trading terminal built around one Alpaca account, virtual strategy sleeves, auditable ledgers, deterministic risk controls, AI-assisted planning, backtesting, and production operations. The product language is dark fantasy; the implementation underneath is boring on purpose: FastAPI, Python services, JSON/JSONL ledgers, Prefect schedules, Alpaca adapters, and Wards that can say no.

What it solves

The design problem was not how to make a trading app look unusual. It was how to give an operator a system that can combine strategy construction, market research, AI planning, risk review, broker execution, and recovery without hiding who owns a position or why an action happened.

I designed the UI and application model around separation of authority: Rituals express desired exposure, Watchers detect Omens, the Oracle and AI Mind propose plans, Wards enforce deterministic law, Manifestation converts approved targets into broker-safe orders, and Chronicle records the chain.

Alpaca Account Vault Ritual Sigil Glyphs Omens / Runes Wards Castings / Bindings

Operator UI

A serious trading terminal wearing dark fantasy armor.

Vault Aether reserve cash, equity, buying power, owner splits

Rituals Strategy sleeves fund, start, stop, sync, release

Oracle / Mind typed AITradingPlan thesis, invalidation, confidence, exit logic

Wards deterministic veto freshness, exposure, cash, High Seal

Vault

One Alpaca account becomes a portfolio view with Aether, broker snapshots, allocation ledgers, positions, owner splits, and PnL attribution.

Rituals

Strategies are virtual sleeves with allocation, lifecycle state, holdings, watcher context, Oracle policy, and Manifestation targets.

Terminal

A read-only operator console exposes status, cycles, logs, Ritual books, broker snapshots, decisions, and stored files.

Backtesting

Divinations test swing, intraday, strategy-block, cross-asset, and watcher-specific ideas before they can approach execution.

Oracle

The Oracle researches, summarizes, proposes, and explains. It can route market evidence into structured plans, but it cannot call the broker, claim execution happened, or bypass deterministic Wards.

Provider fallback, prompt budgets, circuit breakers, and confidence thresholds.
Market research from configured search and macro sources.
Typed AITradingPlan output instead of raw buy or sell commands.

Watchers

Watchers are scouts. They observe data health, market regime, volatility, macro liquidity, theme risk, crypto state, and strategy-specific Omens, then wake the Mind or inform dashboard context.

Theme watchers for miners, black soil, nuclear, AI grid, gold, crypto, and defense.
Runtime intervals, stale-event routing, and stream health checks.
No direct trading authority.

Wards

Wards are final law. They enforce position caps, turnover caps, minimum notional, cash buffer, context freshness, cross-sleeve exposure, regular-hours rules, broker preflight, idempotency, and the High Seal.

AI confidence cannot override deterministic checks.
Broker orders require the execution target and High Seal to be valid.
Blocked plans remain auditable instead of disappearing.

Planning pipeline

AI can scout and plan, but execution stays typed, clipped, and reconciled.

Market data Watchers emit Omens Research adds evidence AI Mind plans Wards validate Planner creates intents Broker or SimBroker executes Chronicle records

Strategy construction

Sigils - composed strategy trees made from deterministic blocks.
Glyphs - reusable conditions, filters, weights, branches, and asset nodes.
Runes - computed metrics such as SMA, RSI, VWAP, volatility, and drawdown.
Omens - market, watcher, or context events that influence planning.
Rituals - deployable sleeves that turn strategy output into target books.

Execution safety

Manifestation - compares desired Ritual books to actual broker positions.
High Seal - kill switch that gates paper and live broker orders.
Ownership ledger - records opened, closed, and migrated position ownership events.
Reconciliation - checks broker state after execution and supports recovery.
Chronicle - preserves decisions, Wards, orders, ownership, and outcomes.

Ritual review surface Each sleeve carries its own thesis, broker reconciliation state, PnL, and backtest path instead of becoming anonymous account-level exposure.

Backtesting and Divination

The backtesting surface tests Rituals before Manifestation. It supports swing, intraday, strategy-block, cross-asset, and watcher-specific simulations with metrics like CAGR, max drawdown, annualized volatility, turnover, time under water, PnL, and benchmark comparison.

The strongest design choice is that backtests reuse live-ish planning components where practical. Divinations are not a toy path beside production; they are a safer rehearsal for the same constraints the operator will face later.

Backtesting lab Divinations keep strategy experiments close to the same broker, Ward, and benchmark constraints used by the live operator surface.

Why it matters as a product system

Glyph keeps the mythic language where it helps the operator reason about roles, but the product is built around practical guarantees: who owns capital, what strategy wants exposure, which risk control blocked a plan, whether the broker state reconciled, and where proceeds go after harvest or deallocation.

What I built toward

A UI where aggressive autonomous planning can exist beside conservative execution law. The Oracle can be ambitious, Watchers can scout constantly, and Rituals can express strategy, but Wards, High Seal, Manifestation, ownership, reconciliation, and Chronicle keep the system inspectable.