Research as a Code

• 

  Finished dynamic strategy UDR deep research top

• 

  Finished dynamic strategy UDR deep research mid

We built the system by integrating several cutting-edge technologies, guided by the principle of "Infrastructure as a Code" (IaaC) for deployment.

• Infrastructure (IaaC): We used the awslabs/data-on-eks Terraform blueprint as a massive accelerator. This single command provisioned our entire foundation:

  • An AWS EKS cluster to host all our microservices.
  • Karpenter for auto-scaling GPU nodes (like g5.xlarge) on-demand.
  • The NVIDIA GPU Operator to automatically configure drivers on new nodes.

• Core Models (NVIDIA NIMs): Once the cluster was up, we deployed the hackathon-mandated NVIDIA NIMs using their Helm charts. This gave us internal, high-performance endpoints for:

  • Reasoning: llama-3.1-nemotron-nano-8b-v1
  • Embeddings: text-embedding-nim (using arctic-embed-l)

• Agent Backend (Python):

  • We started with the NVIDIA AI-Q Research Assistant blueprint, which provided a containerized FastAPI app.
  • We used LangGraph (from the NeMo Agent Toolkit) to define the agent's stateful flow (AgentState) and nodes (planner, tool execution, final report).
  • Our key innovation was wrapping the core logic of the NVIDIA UDR prototype into a single Python function (execute_dynamic_strategy) and registering it as a tool within the LangGraph.
  • We used the copilotkit Python SDK to add a single /copilotkit endpoint to our FastAPI app, which automatically handles streaming the AgentState to the frontend.

• Agent Frontend (React):

  • We built a simple React/Next.js frontend.
  • We used the @copilotkit/react-core library, specifically the useCoAgentStateRender hook. This hook subscribes to the backend's state stream.
  • We wrote a simple render function to map the logs array from our AgentState object into a list on the UI, creating the real-time visualization of the agent's internal flow.

• Deployment: We containerized our custom FastAPI/LangGraph agent, pushed it to ECR, and deployed it to our EKS cluster using a standard Kubernetes Deployment YAML. We configured it to use Kubernetes' internal DNS (e.g., http://nemotron-nano-service.nim.svc.cluster.local) to communicate with the NIMs with zero latency.

Built With

• ag-ui
• ai-q
• amazon-web-services
• claude
• copilpotkit
• deep-research
• deepresearch
• eks
• gemini
• helm
• langchain
• langgraph
• nemotron
• nim
• nvidia
• rag
• terraform
• udr