A trading system is the sum of the assumptions it refuses to violate. Kyoto is built on three.
Implied volatility on Indian indices reliably trades at a premium to realised volatility. The platform is biased — by design — toward systematically selling that premium through defined-risk structures, not toward predicting direction.
A model that can place an order is not a trading system. A model that cannot place an order without passing eighteen SEBI-aware pre-trade checks — ELM, calendar-spread margin, illiquid-strike guards, cost hurdle, vega event warning — is.
Nine narrow servers, each fluent in one concern, expose tools through MCP. The language model orchestrates. Every server is independently deployed, individually testable, and owned by a single phase of the trade lifecycle.
Read the diagram top to bottom: data, then decision, then survival, then edge. Servers within a phase converse through Redis Streams; events between phases follow a small set of standardised contracts.
Live REST surface against Upstox v3 — get_spot_price (5s cache), get_option_chain (30s), get_historical_candles (1h, v3 historical), get_futures_data (5s). WebSocket stream_quotes scheduled in chunk 1.3b.
Seven tools live — calc_iv (Newton-Raphson + Brent, matched broker IV within 0.5%), get_greeks, get_skew, get_term_structure, get_vol_regime (ATM-IV proxy), get_iv_rank / get_iv_percentile (return null until history ingestion lands).
Open-interest concentration, put/call ratio, OI interpolation between NSE three-minute snapshots.
Constructs defined-risk structures, models payoff at expiry and intra-day, returns Pydantic-typed proposals.
The gate. Eighteen pre-trade checks, delta-based OI position limits, settled-cash verification.
Single dispatch surface for three executors. Live mode requires explicit human confirmation per trade.
Eight-year compliance audit trail; per-trade P&L attribution; structured for backtester replay.
Watches institutional newsflow; emits regime-change and event-alert contracts that gate risk and vol.
Replays journaled decisions against historical surfaces; the gate strategies must clear before paper, then live.
Every order — paper or live — flows through the same seven stages. The Risk Manager sits in the middle as a hard gate: a rejected proposal never reaches the broker.
Spot, depth, candles and option chain pulled from Upstox; cached at 5s / 30s.
IV surface fit, rank, term spread, full Greeks and pricing models.
PCR, max pain, OI clusters; interpolation across NSE 3-min snapshots.
A defined-risk structure is proposed — typed, scored, costed.
18 SEBI-aware checks. ELM, settled cash, OI limits, cost hurdle, kill-switch state.
Demo (instant), Paper (sim against live tape), or Live (human confirms, broker routes).
Every fill, slippage, fee and decision rationale persisted for eight years.
★ The gate is non-negotiable. A strategy that cannot pass Risk — for instance, an ITM short on expiry day where projected STT exceeds 30% of premium collected — is rejected at stage five and never reaches an executor. The same gate runs in paper and live; only the destination changes.
Every server speaks MCP. Every server returns the same response shape: a one-line summary, structured data, freshness/source context, and a list of related tools the model might call next.
The single source of truth for spot, depth, option chains and historical candles. Wraps Upstox v2 with circuit-breakers and short TTL caches; emits live tick streams the rest of the platform subscribes to.
Black-Scholes & QuantLib-backed pricing, full Greeks, IV rank and percentile, term-structure spread between current-week and next-month, surface fitting. The desk's calculator.
Concentration heatmaps, put/call ratio, change-in-OI clustering, max-pain. Synthetically interpolates OI between the NSE three-minute snapshots when confidence permits.
Constructs verticals, iron condors, butterflies, ratio spreads, calendars. Returns typed proposals with payoff models, breakevens, expected-value, and a costed round-trip estimate.
Eighteen pre-trade checks aware of SEBI rules: ELM on expiry, calendar-spread margin nullification, settled-cash verification, illiquid-strike guards, vega-event warnings, delta-based OI limits.
Single MCP surface dispatching to DemoExecutor, PaperExecutor, or LiveExecutor by TRADING_MODE. Live requires explicit human confirmation per trade. Reconciliation every 30–60 seconds against Upstox.
Every proposal, gate decision, fill, slippage and fee is persisted with eight-year retention to satisfy SEBI compliance — and structured so the Backtester can replay the exact decision context.
Monitors institutional newsflow against a YAML-driven keyword → risk-multiplier map. Emits regime-change and event-alert contracts the Risk Manager and Vol Engine subscribe to.
DuckDB-backed OLAP over historical chains and journal events. Replays past decisions against historical IV surfaces; strategies must clear backtest before paper, paper before live.
The Trade Executor is the only path to a fill. The same MCP tools — place_strategy, modify_order, cancel_all, square_off_all — operate identically across modes; only the executor behind them changes.
For development & integration tests.
Realistic simulation against live tape.
Real orders, real money, real consequences.
Each layer chosen for one reason. Async wherever the network is; typed wherever the data is; containerised wherever it runs.
SEBI's August-2025 framework around algorithmic trading — static IPs, registered strategies, per-broker order-rate limits — is not a deployment detail. It shapes the architecture.
ELM (+2% on expiry day), calendar-spread margin nullification, settled-cash verification, delta-based OI position limits, illiquid-strike guards, cost hurdle, vega-event warning.
Every proposal, gate decision, fill, modification and rationale persisted to a TimescaleDB hypertable with referenced SEBI circulars in sebi_regulatory_params.
Live trades require explicit operator confirmation within a configurable window. Default 60 seconds — silence is rejection. Confirmations are journaled.
emergency_stop() squares off every position and blocks new orders. Manual and automatic (drawdown, consecutive losses, broker API faults). Re-enable is explicit.
An earlier draft of this design carried primary-and-fallback broker logic. Indian KYC, SEBI algo registration, static-IP allow-listing, and per-broker SPAN margin allocations make true cross-broker failover infeasible at retail scale. So Kyoto uses Upstox v2 alone — and on outage, the kill switch fires and the operator intervenes.
The platform graduates by phase. Each phase ships a coherent slice of the trade lifecycle and is gated by tests and a working integration with the next phase down.
Market Data + Volatility Engine shipping eleven tools with Redis caching. 153 unit + 18 live end-to-end tests green; calc_iv matched live broker IV inside half a percent. WebSocket streaming, instrument-master ingestion, IV history and India-VIX feed tracked as 1.3b / 1.4b / 1.7b / 1.7c follow-ups — none gate Phase 1.
OI Analysis + Strategy Builder. Read what institutions are doing through OI snapshots and translate a view into a typed, costed structure with cost-adjusted POP.
Risk Manager + Trade Executor + Journal. The gate, the dispatcher, the memory. The first phase capable of touching real money.
News Sentinel + Backtester. Adaptation in real time and learning from the past. Strategies graduate through this gate.