Inspiration
I've always been drawn to financial markets watching prices move, trying to understand why stocks behave the way they do. But I never had the intuition to actually trade. The split-second decisions, the emotional discipline, knowing when to buy and when to walk away that never came naturally to me.
What did come naturally was math and code.
So instead of trying to become a trader, I asked a different question: what if I could build something that does the science-based reasoning I could never do myself? Something that studies the market continuously, finds patterns in the noise.
What It Does
Quantara NN is a fully autonomous AI trading agent that studies five stocks continuously using live minute-level market data.
At startup it pulls over 100,000 minute candles per ticker from Alpaca and trains a neural network on each one. Then it enters a continuous loop every 60 seconds it fetches live data, runs four rule-based strategies, generates a neural network signal, and passes everything to NVIDIA Nemotron Super 120B to reason over.
Two specialized Nemotron agents collaborate on every decision. Nemotron Super 120B reasons over technical indicators, strategy performance, neural network signals, and prior decision history. Nemotron Nano 30B translates that technical reasoning into plain English anyone can understand.
Every decision is logged to a persistent SQLite memory layer. Every trade executes through Alpaca paper trading. Everything is visible on a live Streamlit dashboard running on NVIDIA NemoClaw via Brev.
How We Built It
Data pipeline: Live minute candles from yfinance. Historical training data from Alpaca 100,000+ rows per ticker. 14 engineered features including EMA relationships, RSI, MACD, Bollinger Bands, momentum across multiple lookback windows, and volatility measures.
Strategy layer: Four rule-based strategies backtested every cycle and scored with:
$$\text{Score} = 0.5 \cdot \text{Sharpe} + 0.3 \cdot (1 - |\text{MaxDD}|) + 0.2 \cdot \text{Return}$$
The highest scoring strategy generates the rule-based signal for current market conditions.
Neural network: A Multi-Layer Perceptron trained on historical minute data. Input: 14 engineered features. Architecture: $64 \rightarrow 32 \rightarrow 1$ neuron. Output: BUY/SELL with confidence percentage.
Multi-agent reasoning: Nemotron Super 120B handles deep quantitative reasoning. Nemotron Nano 30B handles communication. Two models, two jobs, coordinated every cycle.
Persistent memory: SQLite stores every decision, market state, and portfolio summary. Nemotron reads prior decisions each cycle the longer it runs, the more context it has.
Security: Deployed on NVIDIA NemoClaw via Brev. Policy restricts network access to five trusted domains only, filesystem to project folders, and blocks all shell execution.
Challenges We Ran Into
Data at scale: yfinance caps minute data at 7 days not enough to train a neural network meaningfully. Integrating Alpaca and solving a 15-minute API delay issue on the free tier took significant debugging before the pipeline worked end to end.
Neural network viability: On daily data $\approx 625$ training samples meant the MLP couldn't outperform simpler models. Switching to minute candles gave us 100,000+ samples and made deep learning viable but required rebuilding the entire data pipeline from scratch in under 24 hours.
Conflicting signals: Rule-based strategies and the neural network frequently disagree. Designing the resolution logic when to trust the MLP, when to defer to rules, and how to let Nemotron reason over the conflict was the core architectural challenge.
Autonomous execution: No human in the loop means every edge case has to be handled in code. No shorting positions you don't own. No buying what you already hold. Risk guardrails that reject low-confidence signals before any capital is deployed.
Accomplishments That We're Proud Of
A fully autonomous agent that runs continuously without any human input across five tickers simultaneously.
Two specialized Nemotron models collaborating one reasoning, one explaining as a genuine multi-agent system.
Persistent memory that actually influences Nemotron's reasoning over time the agent genuinely gets smarter the longer it runs.
Real paper trades executing through Alpaca with risk guardrails preventing dangerous actions.
A live dashboard showing signal cards, Nemotron reasoning, plain English explanations, and order history updating in real time.
Built entirely from scratch in 24 hours by a team of two.
What We Learned
Minute-level data changes everything for ML in finance. The same features that underperformed on daily data produced meaningful signals at minute resolution.
Market regime matters more than any single indicator. RSI at 24 means something completely different in a trending market versus a mean-reverting one.
Nemotron's reasoning capabilities are genuinely impressive for structured financial decisions it consistently made defensible calls when signals conflicted. Nemotron carries a more defensive style trading disposition.
Autonomous systems need guardrails at every layer not just security guardrails but logical ones too.
The hardest part of a 24-hour hackathon isn't the code. It's deciding what not to build.
What's Next for Quantara NN
Periodic MLP retraining: The model currently trains once at startup. Adding scheduled retraining on accumulated data would make it genuinely improve over time as it sees more market regimes.
Reinforcement learning: Rather than supervised classification, train the agent on actual trade outcomes reward profitable decisions, penalize losses. True adaptive behavior.
Full NemoClaw enforcement: The security policy is written and the architecture is correct. Wiring up full OpenShell enforcement so blocked actions are demonstrably logged would complete the security story.
Expanded universe: More tickers, more asset classes, more timeframes. The architecture is designed to scale adding a new ticker is a one-line config change.
Options and derivatives: The same indicator framework applies to options flow analysis a natural extension for a more sophisticated trading strategy.
Log in or sign up for Devpost to join the conversation.