-
POSTER
What inspired us
- We were inspired by how useful agents are in operational contexts and the rigor by which Holistic AI has investigated this.
How you built you project
- We took an iterative approach, first debating use cases and how this would be relevant in a financial context before creating a single-observer system. When that worked, we moved on to making this robust by adding a second observer on a different element of the system, the observer memory bank, the interpreter agent, and the logging to facilitate FCA audit requests.
The challenges we faced
Agents are vulnerable at many steps - we focused on a single type of vulnerability (prompt injections) across the agentic system.
What we learned
- Layered security, using varied models is effective.
- You don't need LLMs on everything! Think about implications for latency, robustness.
What's next for Fraud-Y
- Expand known patterns detected by Observer 1;
- Use reinforcement learning to evolve the interpreter agent;
- Detailed ablation study on each component of the system;
- Consider advanced audit logs and connecting observer-like components to traces directly.
VIDEO
Built With
- cryptography
- fastapi
- langchain
- langfuse
- langgraph
- numpy
- openai-gpt-5
- pandas
- pydantic
- python-3.11
- pytorch
- scikit-learn
- sentence-transformers
- sqlite
- streamlit
- uvicorn