Visheshak

Visheshak — विशेषक, "the discerner" Most AI trading demos hand an LLM the keys and hope. Visheshak bets the opposite way: LLMs are good at reading a market regime and terrible at being trusted with order tickets. So it's built as a separation of powers: The Scout (code) builds a market brief from live Alpaca data and constructs a menu of fully-priced, defined-risk candidates from real option chains — credit spreads and iron condors on SPY/QQQ, momentum debit verticals on mega-caps. The Strategist (LLM) reads the brief plus a news digest and maps regime to playbook: uptrend → put credit spread, downtrend → call credit spread, rangebound → iron condor, unclear → abstain. It can only select candidate IDs from the menu — never invent a strike, price, or quantity. Bad output, API failure, or model overload (absorbed by a multi-model fallback chain) all resolve to: abstain. The Risk Officer (pure code, no LLM) holds absolute veto: 1% max loss per structure, 10% total open risk, position caps, daily-loss halt, an equity kill floor, macro-event blackouts, and a "news storm" gate. News is strictly subtractive — it can veto a trade, never create one. The Executor places atomic multi-leg limit orders; a 5-minute Monitor manages profit-takes, stops, pin-risk closes, and a hard flatten deadline. The design test for every prompt rule: "what if the LLM ignores this?" The answer must be "a worse trade inside the limits" or "no trade" — never "money at risk beyond limits." Every decision, veto, order, and exit is journaled append-only, fully auditable. A live indigo-and-saffron dashboard shows equity, day P&L, and active trades straight from the broker. The whole system runs on a free cloud VM with one-command deploys and rebuilds its position state from the broker after any restart — it never trades blind. Raw REST + pydantic, gated by 28 offline failure-injection checks and a synthetic end-to-end test.