Agentic
Market
Intelligence
Most signals get killed before you see them. A 6-bucket convergence gate requires independent agreement across trend, momentum, structure, pattern, institutional flow, and regime context. Survivors are interrogated by a 4-agent AI swarm. What comes out is a thesis — entry, target, stop, and the case behind it.
How It Works
From Price to Signal
01
Raw Data
Every bar. Every instrument.
Institutional tick + bar ingestion across 24 instruments and 6 timeframes. Provider-agnostic bus. Nothing is sampled.
02
Synthesis
Every layer. Every bar.
Indicators, composite events, structure, regime context — computed in dependency order, in a single in-process pass. No polling, no shortcuts.
03
Convergence Gate
Most signals die here.
Six independent evidence buckets — Trend, Momentum, Structure, Pattern, Institutional, Regime — must converge. 3 of 6 required. Score threshold enforced. One strong bucket can't override the rest.
04
AI Interrogation
Survivors get challenged.
Four agents attack each signal before it's published: Skeptic, Correlation, Regime Coherence, Counterfactual. The goal is to find reasons it's wrong.
05
Actionable Signal
Entry. Thesis. Evidence.
Entry, target, stop, regime context, narrative rationale, confidence score. You know why the signal exists and what would have to change for it to fail.
AI Swarm
Adversarial Interrogation
Before a signal reaches you, four agents attack it. Each has a specific challenge. The goal isn't narrative generation — it's adversarial interrogation. Survivors carry a full case.
Skeptic
Finds the strongest case for why this setup fails.
Attacks the entry thesis directly — checks for overextension, adverse regime conditions, recent failure modes for this setup type, and evidence the move is already over.
Correlation
Checks whether macro context supports the thesis.
Cross-asset consistency check — does DXY direction, yield curve posture, VIX regime, and correlated instrument behavior align with the directional call? Contradictions are flagged.
Regime Coherence
Asks if the market regime is right for this setup type.
Trend setups in ranging regimes and mean-reversion setups in strong trends are structurally misaligned. This agent validates HMM regime, session context, and volatility posture match the setup's required conditions.
Counterfactual
What would have to be true for this to be wrong?
Constructs the conditions under which the signal fails — specific price levels, regime shifts, catalyst events — so you enter with a falsification condition, not just a target.
Sample output
Bullish 5m ES setup: trend-regime aligned, SMC confluence confirmed. FVG entry zone 5236–5238, target 5258 (R2.3), stop 5229. Skeptic: no structural concern. Correlation: DXY soft, supportive. Regime coherence: HMM trend-mode, session in NY overlap — valid. Counterfactual: invalidated below 5229 or on VIX spike > 18.
eAI · Evolutionary AI
Agents That Evolve
The swarm isn't static. eAI breeds new agents through evolutionary operators — each generation searching for alpha the current swarm doesn't see. Agents that prove edge get promoted. Agents that don't get demoted. The population evolves.
Agent Genome
Every agent carries a genome — a prompt and configuration parameter set that defines its behavior. The genome is what evolves. Agents aren't hand-tuned; they're bred.
Breeding
Three reproductive operators generate candidates: random mutation for exploration, recombination to combine fit parents, and LLM-directed mutation for targeted search. Each generation searches for alpha the current swarm doesn't see.
Death & Failure Archive
Agents that don't prove edge are demoted. Failures aren't discarded — they're frozen in a gene bank. Failed genomes encode what doesn't work, preventing the population from rediscovering dead ends.
Promotion
New agents start in shadow mode, invisible to production. Promotion requires statistical proof: n ≥ 100 signals, positive expected PnL at the lower confidence bound. The bar is high by design.
4 Layers of Intelligence
I1 → I8 · each layer builds on the layer belowLayer 01
Data Foundation
Ingestion · Bar Building · Stream Distribution
- Institutional real-time tick + bar ingestion
- 1m → 5m / 15m / 1h / 4h / 1d aggregation
- Redpanda stream distribution
- Futures, forex, crypto · 6 timeframes
< 10ms pipeline latency · feed-provider bound, not processing bound
Layer 02
Mathematical Intelligence
Indicators · Structure · Context
- I1: RSI, MACD, ATR, VWAP, OBV, Supertrend, Bollinger Bands, CVD, OFI
- I2: MACD/RSI/stoch events, exhaustion score, momentum acceleration
- I3: Swing H/L, S/R, anchored VWAP, Fibonacci zones, market profile
- I4: GARCH, Kalman, HMM, session context, VIX regime, cross-asset
RSI 67.4 · MACD bullish crossover · GARCH: vol elevated · regime: trend
Layer 03
Pattern Intelligence
Patterns · SMC · Confluence · Signals
- I5: RSI/CMF/MACD divergence, squeeze, chart patterns
- I6: BOS/CHoCH, FVG, order blocks, killzones, AMD cycles, CTF confluence
- I7: Setup plugins — trend, mean-reversion, SMC, session extremes
- CIS: 6-bucket convergence gate · score ±0.35 · 3 of 6 required
BOS confirmed · unfilled FVG 5235–5238 · CIS +0.71 · CHoCHReversal fired
Layer 04
AI Intelligence
Adversarial swarm · eAI · Self-generating alpha
- 4-agent swarm: Skeptic, Correlation, Regime Coherence, Counterfactual — runs on every signal
- eAI breeds new agents: genomes evolve via mutation, recombination, and LLM-directed search
- New agents discover new alpha — the system expands its own edge without manual intervention
- Shadow incubation gates every agent: statistical proof required before production influence
"Bullish 5m ES setup: trend + SMC confluence. FVG entry 5236, target 5258, stop 5229."
Design
Platform Design
Event-Native Architecture
Decoupled by Design
Every stage publishes to a durable stream and consumes from it. Producers and consumers never call each other directly. A new subscriber — alert engine, execution system, ML scorer — connects to existing streams without touching the producers. Extension is additive.
Signal Lifecycle Tracking
8-Class Outcome Model
Every signal is a hypothesis. The system tracks it from emission through activation and exit, recording MAE, MFE, bars-in-trade, and outcome class. Eight outcome labels — from never_activated to target_full — create a labeled training set from every market move.
Regime Context
Multi-Model Consensus
HMM state detection, GARCH volatility modeling, Kalman trend filtering, and VIX regime classification run on every bar. Setup plugins declare a regime type; the aggregator suppresses misaligned setups before anything reaches the convergence gate.
Self-Optimizing Feedback Loop
Evidence-Gated Weights
Outcome data writes back. Setup performance multipliers update from verified PnL and Sharpe roll-ups once 30+ samples are available. The pipeline improves from its own track record — no manual tuning required.
Full Audit Trail
Nothing Is a Black Box
Every LLM call, every signal emission, every outcome is logged with full context — agent ID, prompt version, parse rate, token burn, latency. Adaptive routing promotes models that demonstrate statistical edge. The system is inspectable at every layer.
Feature Store
Ground-Truth Persistence
Every bar’s full feature vector is written to a TimescaleDB hypertable — all 132 plugin outputs, tagged with the weight version that produced them. Signal outcomes JOIN back to their originating feature vector. The history is queryable and replayable.
Architecture
Engineering Principles
Dependency-Ordered Execution
DAG · Topological sort · No polling
Every plugin declares its inputs. The engine runs Kahn's topological sort at startup. Execution order is a mathematical property of the dependency graph, not a convention anyone maintains. Cycles hard-crash at startup. Silent corruption is impossible. Adding a plugin means declaring its dependencies; ordering is inferred.
Regimented Signal Validation
CIS · 6-bucket convergence gate · Evidence required
When 5–8 setup plugins fire on the same bar, CIS decides what gets published and whether anything does. Six buckets (Trend, Momentum, Structure, Pattern, Institutional, Regime) must converge: score > ±0.35 AND 3 of 6 must agree on direction. One dominant bucket cannot override the rest. Discipline enforced by the architecture, not by policy.
Data as Ground Truth
Feature store · Outcome tracing · Self-improving
Every signal is tagged with the weight version that produced it. Every outcome (stop hit, target reached, TTL expired) is written back to the feature store with the full signal vector. Nothing is dropped. When outcome data is sufficient, weights update from evidence. The pipeline captures what it produces and learns from what it captured.
Canonical Typed Vocabulary
IntelligenceEvent schema · Stream-native · API-first
Every output at every tier is encoded into a versioned IntelligenceEvent schema and published to the stream bus. Producers publish. Consumers subscribe. No service calls another directly. A new consumer (alert engine, execution system, ML scorer) subscribes to the existing stream without changing the producers. The bus is the API. Extension is additive.
Shadow-First Governance
Shadow registry · Statistical gate · No silent corruption
Every new plugin and agent enters shadow mode at startup — outputs are computed and logged but carry zero production influence. Promotion requires statistical proof: n ≥ 100 signals and a positive bootstrapped CI lower bound on P&L. Demotion is automatic when edge decays. New code cannot corrupt live signals by default; influence must be earned from evidence.
Compute Ignorant of Persistence
DB-ignorant compute · Hot path isolation · Kafka as sink
Compute agents have no database access — by design, not by convention. I1–I7 runs entirely in-memory and publishes to Kafka. WriterAgents consume from Kafka and own all DB writes. When the database slows or goes offline, the hot path continues unaffected. Backpressure from persistence never reaches the pipeline. The two concerns cannot interfere because they share no code path.
Live Trading Signals
Real-time signals across futures, forex, and crypto markets