DataCore Labs

Agentic AI that runs the full investment research loop autonomously: screening, regime detection, signal generation, execution. Live for two years across 2,500+ trades.

Comment

Inspiration

97% of retail traders lose money. They lack the institutional-grade tooling that hedge funds take for granted: quant-grade research, algorithmic strategies, disciplined risk management, and dedicated analysts working 24/7. We built DataCore Labs to close that gap with a multi-agent AI trading system that thinks and acts on behalf of retail users. When the Nebius x NVIDIA Hackathon opened up access to Nemotron models on open GPU infrastructure, we saw an opportunity to plug NVIDIA's open weights directly into our production agent stack, and to prove that frontier-grade agentic reasoning is possible on open, independent infrastructure.

What it does

Users describe their financial goals, risk profile, and capital in plain English. A five-agent research desk — Data, Research, Strategy, Risk, and Execution — coordinates to pull market data, propose strategies, score risk on five dimensions, and execute inside user-defined guardrails. Every trade improves the system through a self-improving memory loop that feeds a weekly fine-tuning pipeline. During the hackathon we integrated NVIDIA Nemotron models served through Nebius Token Factory as the reasoning engine for our Research and Strategy agents, replacing prior closed-model calls for those specific tasks.

How we built it

Our production stack runs on LangGraph for agent orchestration, BAML for structured LLM output, MCP for tooling (56-tool server), and vLLM for open-weight inference. For this hackathon we routed Research and Strategy agents through Nebius Token Factory using Nemotron 3 Ultra for deep reasoning tasks, and Nemotron Nano/Super for faster, cost-efficient extraction and classification. Nebius Serverless Endpoints handled low-latency inference during live evaluation runs. The self-improving loop uses labeled BAML outputs from Nemotron reasoning traces to feed weekly LoRA fine-tuning on our domain data.

Challenges we ran into

  • Latency versus cost trade-offs. Nemotron 3 Ultra gave strong reasoning but slowed our multi-agent handoffs. Routing between Ultra, Super, and Nano per task type was the fix.
  • Structured output reliability across model families. Our BAML schemas had been tuned on closed models. Adapting them to Nemotron required new prompt engineering and stricter guardrail patterns.
  • Bridging our production system to open infrastructure. Our existing agent stack assumed specific closed-model APIs. Abstracting model calls through a routing layer that could switch between Nemotron on Token Factory and our closed-model providers took most of the hackathon.

Accomplishments we're proud of

  • Successfully migrated our Research and Strategy agents to Nemotron on Nebius Token Factory without regressions in decision quality
  • Achieved comparable win-rate performance to our closed-model baseline
  • Built a vendor-neutral routing layer that lets us swap between open-weight and closed models per task, giving us cost and dependency optionality going forward

What we learned

Frontier-grade agentic reasoning on open infrastructure is real. The gap between closed and open models for well-scoped agent tasks (extraction, classification, structured proposal generation) has essentially closed. For deep multi-step reasoning, Nemotron 3 Ultra held its own against the closed alternatives we benchmarked against internally. Open infrastructure isn't a downgrade; it's an option that expands our design space.

What's next for DataCore Labs

  • Expand Nemotron usage across the remaining agents (Data, Risk, Execution)
  • Deploy our full multi-agent stack on Nebius AI Cloud for regulated-grade audit trails
  • Publish the routing layer as an open-source library so other agentic AI builders can adopt the same hybrid pattern
  • Onboard our first paying retail users in Q1 2027

Built With

Share this project:

Updates

Submission history