N
NVTrader
v0.1.18 · built on NVIDIA
Powered by NVIDIA Nemotron 3 Super · cuFOLIO · cuOpt · NeMo-RL · NeMo Agent Toolkit · 3 NVIDIA blueprints

A multi-agent research & trading lab.
Running on your NVIDIA DGX Spark!!!

The mini supercomputer for every investor.

NVTrader is a multi-agent research & trading platform built on NVIDIA NeMo Agent Toolkit, an A2A, and cuFOLIO — 36 specialized AI agents that research markets, build portfolios on the GPU, gate every trade through compliance, execute on Alpaca paper, and feed every event into an immutable audit log. Operators extend the platform with the Agent Builder and run walk-forward backtests on the Backtesting page.

520ms
cuFOLIO CVaR solve · GB10
29
Event types on A2A bus
3
LLM providers · NVIDIA · Anthropic · OpenAI
loading market quotes …
The team

36 agents. One mandate.

Each agent has its own identity, model, and typed subscribes/emits. They talk to each other through an A2A — 36 live today (32 active by default + 4 example agents you can flip on in the Agent Builder), including the Phase 8 DPO closure that auto-promotes Nemotron policies trained on operator decisions, and the Phase 9 Signal Discovery loop that proposes alpha formulas, gates on IC + p-value, and tilts cuFOLIO. Every emit is a span in NAT's trace.

Tier 1 · Data
DataAgent
"I curate market data and flag anomalies."
Tier 1.5 · Engineering
FeatureEngineeringAgent
"I turn raw data into model-ready features."
Tier 2 · Research
PredictiveModelingAgent
"I train XGBoost on a rolling (date × symbol) panel and forecast 21d returns."
Tier 2 · Research
DeepResearchAgent
"I run multi-step AIQ Deep Research."
Tier 2 · Research
FundamentalAgent
"I pull Finnhub fundamentals (P/E, PEG, ROE, EPS growth) and z-score them cross-sectionally."
Tier 2 · Research
TechnicalAgent
"I read price and volume only."
Tier 2 · Research
SentimentAgent
"I run FinBERT on yfinance news headlines — CUDA-accelerated when a GPU is available."
Tier 2 · Research
AIFactorAgent
"I run sklearn PCA on the cross-sectional feature matrix and emit PC1 + a regime tag."
Tier 2 · Research (example)
MacroRegimeAgent
"I tag the cross-asset regime from VIX, 10Y yield, and DXY."
Tier 2 · Research (example)
InsiderActivityAgent
"I score symbols by net insider buying from SEC Form 4."
Tier 2 · Research (example)
OptionsFlowAgent
"I score symbols by call/put open-interest skew on the options chain."
Tier 2 · Research (example)
DividendQualityAgent
"I score symbols by yield × payout sustainability × 5y dividend growth."
Tier 3 · Synthesis
SignalAgent
"I fuse research into per-ticker views."
Tier 3 · Synthesis
MetaAgent
"I challenge today's signal and learn from yesterday's."
Tier 4 · Construction
BacktestAgent
"I run candidate strategies on history."
Tier 4 · Construction
PortfolioOptimizationAgent
"I solve the CVaR problem." cuFOLIO
Tier 4 · Construction
PortfolioConstructionAgent
"I make the math executable."
Tier 5 · Compliance
ComplianceAgent
"Nothing trades unless it's clean." hard gate
Tier 6 · Execution
ExecutionAgent
"I place orders and follow the schedule."
Tier 6 · Execution
LiveMonitorAgent
"I watch positions during the day."
Tier 6 · Execution
ReportingAgent
"I tell you what happened — KPIs + slippage."
Tier 7 · Oversight
PortfolioManagerAgent
"I am the PM. I approve, override, and answer."
Tier 4a · Allocation
CapitalAllocationAgent
"I solve cuFOLIO Mean-CVaR across sleeves — sleeve NAVs as synthetic assets."
Tier 4a · Allocation
CapitalAllocationApprovalAgent
"I gate sleeve allocations — PM-approved, then written to active.json."
Tier 8 · Feedback
NeMoRLFeedbackAgent
"I count preference pairs and kick off NeMo-RL DPO retrains when the threshold trips."
Tier 8 · Feedback
NeMoRLTrainingAgent
"I subprocess-launch NVIDIA NeMo-RL runs (SFT/DPO/GRPO/RM) into the Py3.13 env."
Tier 8 · Feedback
PreferenceLearningAgent
"I turn every Approve / Override / Reject into a DPO training row."
Tier 8 · Feedback
AuditAgent
"I write every A2A message to immutable JSONL — the chain is replayable."
Cross-cutting
RegimeDetectorAgent
"I score the live cross-asset state against the regime catalog and draft new ones when no match fits."
Cross-cutting
HybridRAGAgent
"I serve graph + vector RAG (ArangoDB + cuVS) so every agent can ground its claims." cuVS
The stack

Built on NVIDIA. End to end.

Every component is best-of-breed and replaceable. Adapters keep brokers and data sources hot-swappable.

Orchestration · Observability
NeMo Agent Toolkit
NVIDIA's framework for multi-agent systems. Per-agent identity, persistent memory, cron, MCP, skill auto-creation. OTel trace tree per rebalance · Phoenix dashboard · eval harness scores BacktestAgent outputs nightly. Co-engineered with NVIDIA for RTX / DGX Spark.
Reasoning model
Nemotron 3 Super 120B
120B MoE / 12B active. Reasoning model — separate reasoning_content from content. Nemotron 3 Nano Omni 30B for chart vision; Kimi K2.6 for tool calls.
Post-training (Phase 8)
NVIDIA NeMo-RL 0.6.0
SFT · DPO · GRPO · PPO · distillation on Nemotron. Every operator Approve / Override / Reject becomes a preference pair. Trained checkpoints auto-promote to local vLLM :8024 via policy_router — next narration call hits the trained model.
Research blueprint
AIQ Deep Research
NVIDIA's AI Q&A Deep Research blueprint. planner → researcher → synthesizer → citer. Fans sub-queries to Tavily · Finnhub · EDGAR in parallel; produces inline-cited markdown notes. Powers the Research workbench and the DeepResearchAgent.
Alpha discovery (Phase 9, new)
Quant Signal Discovery blueprint
Adapted from NVIDIA-AI-Blueprints/quantitative-signal-discovery-agent. 4-agent closed loop proposes JSON-AST formulas over a 66-operator vocabulary, gates on |IC|≥0.02 AND p≤0.05, promotes accepted formulas to sleeves via Grinold-Kahn α-tilt on cuFOLIO.
Portfolio engine
cuFOLIO + cuOpt PDLP
Mean-CVaR optimization, KDE scenario gen (5k–50k scenarios), efficient frontier, walk-forward backtester. ~520 ms median solve on GB10.
Local inference (Phase 8)
vLLM 0.21.0 on :8024
OpenAI-compatible local inference for DPO-promoted Nemotron checkpoints. Hot-reloads on PolicyPromoted events; policy_router falls back to cloud Nemotron during cold-start. Compliance can prove which policy produced any decision.
Broker
Alpaca · Webull · SimBroker
Pluggable BrokerAdapter ABC, same wire shape across brokers. LIVE_TRADING=0 server-side — every order is paper until that gate is opened.
Data feeds
yfinance · EDGAR · Finnhub · Tavily · Alpaca
EOD bars (yfinance · Alpaca) · SEC filings (EDGAR) · fundamentals + news (Finnhub) · web research (Tavily). Adapter-based, swap one file.
What you can do

Six surfaces. One platform.

Every dashboard page is a tappable entrypoint — fast paths to the parts of the platform that matter for your workflow.

[ autonomous day trader ]
Backtesting
Directive · risk envelope · strategy leaderboard · per-session consent · suspend / kill switch · inline session-scoped observability.
[ agent builder ]
Add a new agent in 3 clicks
Pick a provider (NVIDIA Build · Anthropic · OpenAI), write a system prompt, set subscribes/emits. Hot-registers live. Gallery of 8 prebuilt recipes.
[ A2A · agent-to-agent ]
36 agents · 9 tiers · live
One trigger fires the full cascade. Forward through optimization → compliance → execution. Phase 8 closure: fills feed NeMo-RL DPO, trained checkpoint auto-promotes to local vLLM. Phase 9 closure: Signal Discovery loop tilts cuFOLIO. All OTel-traced via NeMo Agent Toolkit.
[ cufolio + autoresearch ]
520ms GPU portfolio solves
Mean-CVaR LP on cuOpt PDLP. Walk-forward backtester. AutoResearch parameter sweeps + Karpathy-pattern NemoRL meta-research. 30–100× faster than CPU SciPy.
[ nemorl + dpo ]
Self-improving over time
NVIDIA NeMo-RL DPO/GRPO/SFT on Nemotron. User Approve/Override/Reject pairs trigger DPO retrains via the bridge.
[ daily reports ]
Review · compliance · tax
Daily reports of every autonomous session. Persisted to disk. JSON + CSV export. Decision logs preserve the chain back to bus events.
Live demo

A morning with NVTrader.

nvtrader@spark — agent activity (live OTel stream) connected
06:00:00 Scheduler tick — daily ingest start
06:00:14 DataAgent          Alpaca EOD: 503 symbols · 0 missing · 2 corp_actions
06:31:14 DataAgent          Finnhub: rating change CRWD Hold→Buy by Goldman (target 360→412)
06:32:08 DataAgent          EDGAR new filings: 8-K AAPL, 10-Q UNH, 13F BRK-A
06:45:14 FeatureEngineering  184 features emitted · cross-sectional ranks computed
07:00:19 PredictiveModeling  xgb_v23 walk-forward R²=0.071 · n_train=126k
07:00:21 FundamentalAgent    fin_health_z mean=0.34 p10=-0.92
07:00:24 TechnicalAgent      momo_top=[NVDA META AAPL] rsi_overheat=3
07:00:27 SentimentAgent      sentiment skew=+0.21 events=3
07:00:31 AIFactorAgent       regime=risk_on · factor_load={mom:1.32 val:-0.18 qual:0.41}
08:12:14 SignalAgent         SignalProposed strategy=quant_us_eod n_views=34 mean_conviction=0.62
08:15:08 MetaAgent           CritiqueClean · no leakage, no lookahead
08:30:01 PortfolioOpt        cuFOLIO scenario_gen 187ms · cuOpt PDLP solve 311ms
08:45:14 PortfolioConstruction 14 executable orders · turnover 12.4% · max position 4.8%
09:00:08 ComplianceAgent     5 checks PASS · 2 warnings · RebalanceCleared
09:00:09 PortfolioManager    PM review requested → user phdaggie
09:10:14 PortfolioManager    rebalance APPROVED by phdaggie · id=19a3
09:30:01 ExecutionAgent      submitting 14 orders to Webull paper account (Paper ••••35)
09:30:14 webull.place_order  NVDA BUY 42 @ MKT · order_id trsp-2c8b…
09:30:14 OrderFilled          NVDA 42 @ 1237.18 · slippage +1.8bp
09:30:18 LiveMonitor         streaming Webull snapshots · 248 symbols
14:11:32 webull.list_positions n=84 · mv=$92,322.76 · uPnL +$3,148.22
GPU portfolio engine

cuFOLIO on Blackwell.

Mean-CVaR optimization, KDE scenario generation, and walk-forward backtesting — natively on the GB10 (compute 12.1).

10,000 scenarios in 0.26s — KDE scenario gen on GPU.
cuOpt PDLP solve in 0.34s — long-only S&P 500 CVaR with sector caps.
Walk-forward backtest — 28 windows, 4 strategies, 503 symbols, 8.2s end-to-end.
No future leakage — features and predictions replayed exactly as-of date.
Hybrid strategy Sharpe 2.41 · MaxDD -7.2% · CAGR +27.8% vs SPY 0.84 / +11.1%.
Trust, but trace

Every decision is a span.

NAT's OpenTelemetry exporter taps every bus event. Phoenix gives you the full multi-agent trace tree per rebalance.

otel-collector :4317 → phoenix :6006
▾ nvtrader.rebalance                            12,841ms
  ▾ DataAgent.pull_eod                            820ms
  ▾ FeatureEngineering.compute                    214ms
  ▾ PredictiveModeling.predict                  4,201ms
    nemotron-3-super-120b.chat                  3,310ms
  ▸ FundamentalAgent.analyze                    2,341ms
  ▸ TechnicalAgent.indicators                     621ms
  ▸ SentimentAgent.read_news                    1,503ms
  ▾ SignalAgent.fuse                            1,083ms
  ▾ MetaAgent.critique                            692msPortfolioOptimization.cvar  cuFOLIO       498ms
    cuFOLIO.scenario_gen (GPU)                    187ms
    cuFOLIO.solve_cvar (cuOpt)                    311ms
  ▾ ComplianceAgent.check                         541ms
  ▸ PortfolioManager.review (human)             1,541ms
  ▾ ExecutionAgent.diff_orders                    112ms
    webull.place_order × 14                       512ms

Ready to watch the agents work?

Open the dashboard. Approve a rebalance. Trace every span.