About Strat AI — Definitive Product Context & Engineering Architecture
Explore the quantitative engineering blueprint, multi-agent AI research loop, Rust binary tick pipeline, and capital preservation philosophy of Strat AI.
Pre-trade risk adjudication engineered with mathematical rigor
Strat AI is a market analysis and pre-trade risk adjudication terminal for the Indian stock market (NSE). Developed by the Trading and Research Wing, it combines multi-agent research evaluation, deterministic risk verification, and streaming glass-box reasoning.
Trading requires objective analysis over emotional reaction
Market analysis is often cluttered by conflicting media headlines and lagging indicators. Strat AI acts as an objective pre-trade research co-pilot, evaluating setups against hard volatility constraints, order flow imbalances, and options concentration before capital is committed.
Pre-trade risk discipline & honest failure
Our philosophy is built around capital protection through deterministic rules: unbypassable 1.5× ATR stop floors, profile reward-to-risk constraints, automatic stand-asides during data conflicts, and transparent reporting when data is unavailable.
The 3 Pillars of Strat AI Intelligence
Built on a stateful multi-agent reasoning engine, linear regression trajectory projection, and dual technical/sentiment signal fusion.
Stateful Multi-Agent Reasoning Engine
A structured multi-agent research pipeline operating across 4 stateful decision modes.
FIND Mode (15-Step Scan)
Evaluates macro trends (1H/4H/1D), VWAP, volume profile Point of Control, VWEPR quadratic curvature, Order Flow Imbalance, and 19 chart patterns (>0.6 confidence).
VERIFY Mode (Risk Audit)
Stress-tests trade setups against strict 1.5× ATR volatility stop floors, minimum R:R constraints (1:1.3 intraday / 1:2 swing), and unleashes a Bear Agent critique.
DEBATE Mode (Consensus)
Spawns competing Bull & Bear AI agents to debate market thesis. A Judge Agent computes weighted conviction, applying a 25-point penalty if the debate remains contested.
QA Mode (Glass-Box Audit)
Interactive plain-language auditing where traders probe the AI's exact reasoning. Committed trade decisions remain immutable during Q&A to preserve auditability.
10-Minute Trajectory Projections (OLS Regression)
Projects forward price trajectories onto 10-minute charts using rolling ordinary least squares regression.
Mathematical Rigor & Lock
- ✓14-Candle Rolling Window: Maintains Ordinary Least Squares (OLS) linear regression across the last 14 closes on 10-minute candles.
- ✓R² Confidence Score: Displays the exact Coefficient of Determination (R²) so traders know how well recent price fits the regression line.
- ✓Timeframe Integrity Lock: Trajectory projections only render on 10-minute charts where calibrated, preventing misleading projections on wrong timeframes.
Fused Conviction Score (1–100) & Capital Guardrail
Synthesizes technical indicators, news sentiment, and breakout anomaly streams into a single relative setup ranking.
Conviction Fusion & The Conflict Rule
The Aggregator engine blends technical momentum at 70% weight and news sentiment at 30%, but inverts that to 30/70 once sentiment conviction exceeds 85, on the reasoning that strong news breaks technical patterns.
Measurement Separated From Reasoning
Built with Rust, Kafka, QuestDB, and Tauri desktop native IPC. Every number the model reads came out of a deterministic function first.
Rust Binary Tick Parser
Connects to Zerodha Kite WebSocket streams, parsing raw binary tick packets in Rust with zero garbage collection pauses.
Dual-Sink Data Pipeline
Simultaneously publishes ticks to Kafka/Redpanda for live AI agent processing and QuestDB for historical time-series storage.
Order Flow & Footprint Rendering
Renders volume profile Point of Control (POC) and tick-level bid/ask footprint imbalances directly on WebGL/Canvas. Below the minimum usable tick count, order flow returns nothing rather than a neutral zero.
7 Foundational Principles of Reliability
Engineered to eliminate AI hallucinations and enforce deterministic pre-trade risk rules.
Honest Failure Over Fabrication
When an API or data source times out, the system marks it as "unavailable" rather than fabricating neutral values. Zero synthetic data.
Unbypassable Hard Risk Rules
Stop losses must be ≥ 1.5× ATR and R:R must clear minimum thresholds (1:1.3 intraday / 1:2 swing). Enforced deterministically in both Rust and Python.
Asymmetric Conflict Rule
When a strongly bearish technical read meets strongly bullish sentiment, the blended conviction is pulled 60% toward neutral rather than averaged. The reasoning loop is separately permitted to conclude HOLD.
Tamper-Evident Recommendation Log
Records every recommendation with tool inputs, prompt hashes, and model IDs in an immutable, append-only store.
Full Glass-Box Transparency
Streams every tool call, data point, and reasoning step live to the screen as it happens. Watch the system evaluate in real time.
Adversarial Self-Critique
Bear Agent critique stress-tests every trade setup against VWAP resistance, option walls, and session chop before outputting research.
Four Focused Workspace Profiles
Each profile carries its own reward-to-risk floor, sidebar, remembered instrument, and agent tool binding, enforced in the store rather than hidden in the interface.
1. Intraday Scalper
1m/5m charts, live L2 bid/ask order book depth, and intraday volatility heatmaps.
2. Swing Trader
Multi-timeframe trend alignment (1H, 4H, 1D, 1W), Fear & Greed gauge, and news sentiment scoring.
3. Investor Mode
Macro indicators (Fed funds, CPI, Treasury, VIX), discipline metrics, and sector trend analysis.
4. F&O Mode
Options chain analytics, Greeks, OI buildup and walls, with the options tool bound to the agent and split view enabled.
Experience quantitative analysis with mathematical rigor
Join Indian traders using Strat AI to evaluate setups, audit risk parameters, and stream glass-box research.