[FEA][Proposal] Community Example: Neuro-Symbolic Agent Benchmark using cuDF (4.5M ops/sec)

[saimahesh8752]

Is your feature request related to a problem? Please describe.
As "Agentic AI" workflows become more common, there is a lack of community examples demonstrating how cudf can be used for high-frequency decision verification (Neuro-Symbolic AI). Most tutorials focus on ETL or clustering, missing the use case of "Runtime Guardrails" for LLM agents.

Describe the solution you'd like
I have developed a benchmark script (OmniGuard) that demonstrates how cudf and cupy can be used to vectorize complex boolean safety rules for AI Agents.

Benchmark Results (NVIDIA T4):

• Throughput: 4.5 Million verifications/sec
• Latency: ~0.2s for 1 Million Agents (Batch)
• Speedup: ~90x faster than standard Python/Pandas loops.

I have verified this on Kaggle using the RAPIDS T4 image.
Live Benchmark: [https://www.kaggle.com/code/saimaheshsandeboina/omniguard-neuro-symbolic-safety-benchmark?scriptVersionId=290644627]

Proposal:
I would like to contribute a cleaned-up version of this notebook to the community examples (or notebooks-contrib) to demonstrate:

1. cudf for boolean masking in Agentic workflows.
2. Integration with Neuro-Symbolic logic.
3. Real-time performance benchmarking on GPU.

Is this something the team would be interested in accepting as a PR?

[btepera]

Hi @saimahesh8752 - thanks for opening this issue and sharing this example! As you mentioned, https://github.com/rapidsai-community/notebooks-contrib/ is probably the best place for this type of contribution. If you want to open a PR there with the cleaned up notebook we can go through it in more detail but here are a few pieces of initial feedback:

• It looks like the speedup being measured is for the boolean masking component of this script. We have some existing community notebooks that contain similar examples using cudf to filter datasets based on different criteria. What is special about the filtering being done in this workflow? Is high throughput boolean masking typically a key bottleneck for this type of workflow?
• Re: benchmarking, it'd be good to have code demonstrating the same operations running on the CPU as well (assuming it's a reasonable runtime) to validate any speedup claims being made.
• More background or details on the use case would be helpful to users not familiar with the workflow. What problem is OmniGuard trying to solve? Are the sample data and dummy rules in this notebook meant to be replaced with a different model at some point, or is this demo more of a postprocessing step after receiving some upstream model output? What about the "MarketSense" decision layer?

I think addressing some of these questions would be a good first step toward contributing an effective demo notebook. Feel free to reach out if you have additional questions, otherwise we can move this discussion to the other repo once you think the notebook is ready.

[saimahesh8752]

Hi @btepera, thanks for the detailed feedback — this is very helpful.

To clarify the intent before opening a PR in notebooks-contrib:

1. Use case and motivation

You’re right that the core operation is boolean masking, and similar patterns already exist in the repo. The goal of this notebook is to contextualize that operation within a high-throughput decision-validation workflow, where large volumes of model outputs must be filtered against multiple deterministic constraints under tight latency requirements. In these scenarios, the rule-evaluation step becomes a dominant bottleneck, which is where GPU-accelerated, vectorized filtering can provide meaningful value.

2. CPU benchmarking

Agreed on the importance of validating performance claims. I’ll include a side-by-side comparison using equivalent pandas logic on CPU and cuDF on GPU, measuring end-to-end execution time for the same dataset and rules to make the comparison explicit.

3. Context: MarketSense and OmniGuard

I’ll expand the notebook introduction to clearly explain the workflow: simulated probabilistic outputs representing an upstream model or LLM (“MarketSense”), followed by a deterministic rule-based validation step (“OmniGuard”). The rules in the notebook are simplified stand-ins intended to demonstrate the post-processing and enforcement pattern rather than a full production policy engine.

I’ll clean up the notebook with this additional context and benchmarking and then open a PR in rapidsai-community/notebooks-contrib for further review.

Thanks again for the guidance.

[saimahesh8752]

Hello [@btepera,]
Update: I have implemented this proposal as a community tutorial based on the feedback here.

The benchmark and documentation are now submitted in the notebooks-contrib repository.

PR Link: (rapidsai-community/notebooks-contrib#391)

Ready for review.
Thank you.