NVTrader reference architecture

36 specialized AI agents on the A2A event bus, end-to-end on a single NVIDIA DGX Spark / GB10. Trading research → cuFOLIO GPU optimization → 10-check validation report (Nemotron-narrated) → compliance-gated execution → realized-compare (backtest ↔ paper ↔ live, one strategy_version_id spine) → post-trade drift analysis → fills feed NeMo-RL DPO retrains → auto-promote to vLLM on localhost:8024 (Phase 8 closure) → next narration call hits the trained policy. Phase 9 — LLM-driven Signal Discovery closes a second flywheel: 4 Nemotron-driven agents (Generator · CodeGenerator · Advisor · Orchestrator) propose JSON-AST formulas over a 66-operator vocabulary, evaluate Mean IC + p-value on real yfinance bars, gate on |IC|≥0.02 AND p≤0.05, and promote accepted formulas to sleeves with Grinold-Kahn α-tilt on cuFOLIO scenarios. Architecture adapted from NVIDIA-AI-Blueprints quantitative-signal-discovery-agent. Full OTel audit on every step (including LLM chain-of-thought via llm.reasoning.* spans).

Layer diagram

The seven-and-a-half layer stack — each layer's responsibility, the NVIDIA-owned components, and how the pieces talk. Scroll inside the frame for the full SVG.

Layers at a glance

End-to-end data flow

One trigger fires the full pipeline. ~40 events across 36 agents in ~5 seconds on GB10 (excluding training runs). The Phase 8 DPO closure adds an asynchronous outer loop that auto-promotes trained checkpoints back into inference. The Phase 9 Signal-Discovery loop runs on-demand from a chat intent ("momentum signals on US tech") and typically completes 3 iterations in 60-120 seconds against build.nvidia.com Nemotron 3 Super.

OPTIONAL OUTER LOOP — multi-strategy capital allocator
  ─────────────────────────────────────────────────────
  MultiSleeveRebalanceRequested
  → CapitalAllocationAgent        (cuFOLIO Mean-CVaR on sleeve NAVs)
  → CapitalAllocationProposed     → PM Agent gates approval
  → CapitalAllocationApproved
  → ComplianceBusAgent (allocator gate: sleeve caps + cross-sleeve concentration + sanity)
  → AllocationCleared / AllocationBlocked
  → CapitalAllocationApprovalAgent → writes data/allocations/active.json
  → CapitalAllocated (terminal; fans out to per-sleeve flow below)

  PER-SLEEVE INNER LOOP
  ─────────────────────
  Scheduler.tick.eod
  → DataAgent              → DataReady
  → FeatureEngineeringAgent → FeaturesReady
  → Technical · Fundamental · Sentiment (parallel) → ResearchComplete × 3
  → SignalAgent            → SignalProposed
  → PortfolioOptimizationAgent (cuFOLIO solve, ~520ms) → RebalanceProposed
  → PortfolioConstructionAgent  ← reads active.json: capital = equity × pct[sleeve]
                               → RebalanceConstructed
  → ComplianceBusAgent     → RebalanceCleared / RebalanceBlocked
                             (cross-sleeve concentration check fires here
                              when active.json has > 1 sleeve)
  → PortfolioManagerBusAgent → RebalanceApproved
  → ExecutionBusAgent      → OrderPlaced (real broker call)
  → LiveMonitorAgent       → OrderFilled

  ⇢ STRESS-REGIME BRANCH (off DataReady)
  DataReady → RegimeDetectorAgent → RegimeMatchProposed / RegimeDraftProposed
            → RegimesChanged → PortfolioOptimizationAgent (synthetic stress
              scenarios injected into cuFOLIO solve · same pipeline)

  ⇢ PHASE 0-6 LIFECYCLE (per strategy_version_id)
  BacktestReport → BenchmarkingAgent (train/val/test sweep, one-shot test)
                 → BenchmarkReport
  BenchmarkReport → ValidationNarrationAgent (Nemotron narrates 10 checks)
                  → ValidationReportReady
                  → ComplianceAgent (strategy-definition gate: restricted list, universe>50)
                  → ComplianceAdvisory
  ValidationReportReady (gate ≥ 7/10) → eligible for paper deploy
  Deploy modal (live broker only): frozen params + 6-required checklist
    → POST /api/lifecycle/deploy → immutable Deployment record
  Paper/live activity → RealizedCompareAgent (backtest ↔ paper ↔ live, joined by strategy_version_id)
                      → RealizedCompareReady (drift signals: slippage_delta_bps, behavior_match)
                      → PostTradeAnalystAgent (5 drift dims, 7-action recommendation, Nemotron narrates)
                      → PostTradeAnalysisReady

  ↺ FEEDBACK · PREFERENCE FLYWHEEL
  RebalanceApproved/Rejected/Override
    → PreferenceLearningAgent → preferences/extract.py (T+5 yfinance backfill)
                              → preferences/trainer.py build_dpo_dataset
                              → outcome-weighted DPO with label inversion
                                 (weight = f(decision, vs_spy_pct sign);
                                  boost when decision & market agree,
                                  invert chosen↔rejected when they contradict)
    → PreferenceRecorded → NeMoRLFeedbackAgent (counts pairs)
                         → TrainNemoRLRequested (algo='dpo')
                         → NeMoRLTrainingAgent (subprocess → Py3.13 env)
  TRL DPO loop (parallel, in-process):
    → LoRA on TinyLlama-1.1B → data/rlhf/adapters/<ts>/
    → preferences/style_adapter.py hot-reloads → Style-match % + AUC + drift badge

  ↺ PHASE 8 — DPO CLOSURE (the flywheel finally spins)
  NeMoRLTrainingComplete (from runner pump thread)
    → PolicyPromotionAgent
       · find_latest_checkpoint(run_id, algo)         results/<algo>/<run_id>/step_N
       · ensure_hf_checkpoint(step_dir)               DCP→HF via NeMo-RL converter subprocess
       · parse_final_metrics(log_path)                val_loss · val_reward · val_accuracy
       · baseline_metrics_for(decision_type)          prior active OR fallback floor
       · gate_passes(eval, baseline)                  val_loss↓ OR val_reward↑
       · register_candidate + promote                 candidate → active
    → PolicyPromoted → InferenceServerAgent.reload(checkpoint_path)
                     → vLLM hot-reload on localhost:8024 (--served-model-name nvtrader-local-<run_id>)
    → LocalInferenceReloadStarted
  Next narration call: policy_router.call_llm(decision_type, ...)
       · active policy exists + vLLM healthy → route to local
       · else → build.nvidia.com Nemotron 3 Super (cloud fallback)
       · RoutedReply.policy_id stamped on response → audit captures who narrated

  ↺ PHASE 9 — LLM-DRIVEN SIGNAL DISCOVERY
  (Adapted from NVIDIA-AI-Blueprints quantitative-signal-discovery-agent)
  SignalDiscoveryRequested(intent="momentum on US tech mega-caps")
    → SignalDiscoveryOrchestratorAgent
       · (if universe omitted) universe_resolution LLM picks sector slice
         from a 13-slice catalog → UniverseResolved
       · for iter in 1..max_iterations:
            SignalGeneratorAgent   (Nemotron 3 Super, temp=0.8)
              · sees operator vocabulary (66 ops) + intent + best-so-far + last_feedback
              · emits SignalCandidatesGenerated { signals: [JSON AST, ...] }
            SignalCodeGeneratorAgent
              · validate AST against strict whitelist (no exec)
              · compile to vectorized callable via OPERATOR_REGISTRY
              · emits SignalCodeCompiled
            evaluator (deterministic)
              · pull real OHLCV via yfinance (window matches user dates exactly)
              · Spearman rank-IC + p-value + IR + decay + quintile spread
              · emits SignalEvaluated × N
            acceptance gate (|IC|≥0.02 AND p≤0.05)
              · pass → SignalAccepted → save data/discovery/signals/<id>.json
              · fail → SignalRejected
            (if all rejected and not last iter)
              OptimizationAdvisorAgent (Nemotron 3 Super, temp=0.5)
                · operator-grade critique: "Try TS_Rank instead of CS_Rank,
                  add TS_Zscore vol normalization, gate when..."
                · emits OptimizationAdviceGenerated → binds next iter's prompt
       · terminal: SignalDiscoveryComplete { status: accepted|best_effort|failed }
  Promote-to-sleeve: POST /api/discovery/signal/accepted/<id>/promote-to-sleeve
    → configs/sleeves/discovered_*.yaml with signal_kind=discovered_formula
    → run_fold loads sleeve → discovery/alpha_overlay.compute_alpha_tilt
       (Grinold-Kahn α = z · σ · IC_target)
    → tilt_prices(window_prices, α_daily)   # Black-Litterman view at data layer
    → cuFOLIO.generate_scenarios on tilted prices  → real signal drives weights
    → FoldResult.signal_overlay records formula + n_rebalances_tilted

  Reasoning trace: every Nemotron call in Phases 8/9 emits llm.reasoning.<role>
    spans into data/traces/spans.jsonl with chain-of-thought (reasoning_content),
    token usage, policy_id provenance. Surfaced on /observability.html and
    inline on the Signal-Discovery iteration timeline.

  every meaningful event (60+ subscribed) → AuditAgent → data/audit/bus_events.jsonl
                                          (tagged with strategy_version_id, strategy_id, mode)

The outer loop is operator-triggered. When no allocation has been approved, the inner loop runs in single-sleeve mode (each sleeve sizes against full account equity), preserving today's behavior. See the multi-strategy capital allocation skill card for full operational detail.

Full event vocabulary on the Event vocabulary page; full agent roster on the Agent roster page.

NVIDIA components map

See also

• Full technical brief — the CEO/partner-ready long form.
• A2A event bus engine docs — the substrate that ties it all together.
• Standalone diagram ↗ — full-screen, screenshot-ready for slides.