A five-agent SPX 0DTE options pipeline running on sovereign dual-DGX GPU infrastructure — human-in-the-loop by default, every decision traceable end-to-end.
Built AI-first: local LLMs reason inside a deterministic risk cage designed so no model can bypass the gates. Today the system trades in a broker-dealer paper environment under human-in-the-loop control while a structured fund-launch program takes shape around it.
Market signal to executed order passes through five specialized AI agents — each with a narrow mandate, frozen data contracts between them, and a single trace ID stitching the full causal chain from signal to fill.
Scans market structure for actionable dislocations and volatility events.
Designs defined-risk options structures from the signal, greeks-first.
Independently validates every proposal against risk rules before it advances.
Places human-approved orders through the broker layer — limit-only entries.
Post-trade review and news correlation feed back into the next cycle.
Two brains, two latency budgets
On-prem, no per-token cost in the hot loop
Every order explains itself
The design principle: LLMs perform bounded reasoning inside a deterministic risk cage — designed so no model can bypass the gates, with the human holding the last word on every order.
By default, every trade memo requires operator approval in a dedicated control UI before any order is transmitted. A tightly bounded autonomous-approval pilot (single-contract size, hard per-order loss cap, small daily order limit, symbol whitelist) now runs in the paper environment under staffed operator monitoring — human approval continues in parallel for everything outside the pilot's bounds.
Operational — paper tradingEntry orders must be limit orders — enforced at three independent layers (prompt, data model, execution gateway). Market orders are reserved for exits and protective liquidation.
Operational — paper tradingA live oversight file is read before every proposal cycle. If it says halt, the system halts — a human-writable kill document that outranks every agent.
Operational — paper tradingA two-step, armed kill-switch in the operator UI — every activation logged, and exercised in a real, logged operational drill in July 2026. Paired with a capital-protector daemon that monitors positions against loss thresholds.
Operational — paper tradingPre-flight checks in every execution gateway refuse malformed orders — a zero or missing quantity can never reach a broker.
Operational — paper tradingPaper and live stacks are fully state-isolated — separate databases, separate credentials. The paper stack is provisioned without live-broker credentials, so it cannot place a live order by design.
Operational — paper tradingA pre-trade margin gate in the executor validates every order against deterministic margin math before submission — including refusing naked-short structures outright. Promoted from staged audit to enforcing in the paper environment after a monitored soak.
Operational — paper tradingPortfolio-level exposure caps (delta, vega, notional) and per-underlying concentration limits are checked before proposals advance and again at execution; risk-reducing closes are exempt by design. Blocks are enforcing in the paper environment, with every decision written to an audit trail.
Operational — paper tradingA deduplication gate fingerprints every candidate at two levels — exact trade and structure — and suppresses repeats before they reach approval. Close and roll intents are never suppressed, so risk-reducing flow always passes.
Operational — paper tradingFill state is reconciled against broker records on a durable order-identity bridge, with a lifecycle writer that pages the operator on ledger-versus-broker divergence — the system's book is reconciled against the broker's, continuously.
Operational — paper tradingSpread submissions are watched for naked-leg exposure; detection and paging are operational, with graduated response rungs (probe, cancel-to-flat, residual close) built and staged behind an arming gate.
Alerting operational — actions stagedAn automated daily-loss halt monitors broker-reported P&L against dual thresholds (percentage of equity and absolute dollars, worse-of-two) and halts new order submissions on breach — trip-tested in a logged drill before enforcement. The fuller tiered ladder (hedge-only mode through liquidation lock-out) remains a design target.
Operational — paper tradingDesigned control: no strategy reaches production without a signed manifest — code hash, out-of-sample metrics, overfitting probability bound, and a paper-trade attestation window.
Design targetThe process is the product: changes to anything safety-relevant follow a fixed ritual designed to make failure boring.
Plan → adversarial review → land inert → audit soak → posture-gated enforcement. Nothing safety-relevant flips on in one step.
Every behavior is verified against a simulated broker and deterministic reasoning stubs before any live LLM or paper-trading test touches it.
Every risk control inherits the data tier of its inputs — automated actions on delayed market data are forbidden by policy.
Seven engineering rules — from limit-only entries to network isolation of the execution path — enforced by tests that must pass on every commit.
Safety-relevant changes pass independent review before code is written and again before merge; the accumulated codebase is periodically swept by parallel reviewer agents whose findings must each survive an adversarial verification pass before a fix ships. Confirmed findings become tracked, individually reviewed fixes.
Beyond the 0DTE pipeline, the platform is growing a second strategy sleeve and the execution machinery to trade at institutional size — each piece labeled by what it is today.
Every fill measured, honestly labeled
A second strategy, built research-first, now wired to the paper broker
Engineered toward institutional-scale intraday capacity — design target
The strategy program is converging on a classic hedged-book shape — a single-name long/short book with option income written against it — rebuilt so that every position is born and dies inside one session. Each sleeve is labeled by what it is today.
Intraday long/short single names, flat by the close
Designed to collect time decay against the book, one session at a time
Tail protection sized to the book's residual beta
Agentic trading does not replace the discipline of a traditional portfolio manager — it encodes it. The techniques below are decades old; what is new is that each one is a gate in software, with a human on the other side of every switch.
Holding period and instrument choice are risk parameters, not afterthoughts
Protection is bought deliberately, at a known cost
The oldest options income technique, compressed to one session
Per-name, per-underlying, and account-wide caps that fail closed
Long/short, covered income, tail hedges — with agents doing the bookkeeping
Half size first, full size after clean sessions
Beside the intraday book, ieBalancer rebalances one account, a household or a batch of accounts: full rebalances, raising or investing cash, and targeted adjustments. Every sell it proposes passes the tax rules set for that client, each order set is validated at the broker-dealer before anything is placed, and a person approves exactly the set that was validated.
ieBalancer applies tax rules; it does not give tax advice. Wash-sale and basis figures are estimates, and the broker-dealer's 1099-B is authoritative. In development: the rules listed here are built and tested on a simulator; harvesting as a scheduled run and option overlays are not yet built.
In developmentIgniteEdge is being developed as the trading engine for a planned private fund pursuing a systematic, hedged intraday program: an equity long/short book, a defined-risk 0DTE index-options income sleeve — maximum loss known at entry — and an index-put hedge, with intended mandates of capped net exposure, bounded hedge cost, and ending each session flat.
A structured launch program is underway alongside the engineering work: legal formation, compliance policies, and an operating spine of independent service providers — fund administrator, auditor, outsourced compliance, counsel, custody, and banking. Every external-facing claim passes a dedicated review gate before it ships, and a hard readiness gate stands before any solicitation begins.
The system currently operates in a broker-dealer paper-trading environment with no client capital. No performance figures are published here — when they exist in distributable form, they will come with full methodology and provenance. Investor materials, if any become available, are provided only through a private review process and only to eligible prospective investors.
If you want to talk agentic trading infrastructure, GPU inference, or risk engineering — reach out.
The engineering is ahead of the paperwork. What the launch program needs next is the non-technical operating spine — and warm referrals beat cold search. If you know good people in any of these areas, an introduction would genuinely help.
To be clear: this is a request for professional referrals and expertise only — not for capital. Nothing on this page is an offer or a solicitation of any investment.