A multi-agent AI fraud detection system in 6 hours

Our team competed in the Reply Mirror AI Agent Challenge. The mission: detect evolving financial fraud in an environment where attackers constantly adapt.
Instead of a single model, we designed a multi-agent system:
- Memory Agent — learns user behaviour patterns over time
- Narrative Agent — interprets communication context (SMS, email)
- Consensus Mediator — aggregates signals across modules
- LLM Judge — makes the final fraud decision
It covers transactions, user profiles, geolocation, messages and audio, with a score-fusion engine, threshold auto-tuning and Langfuse for full decision traceability.
What we learned: fraud detection isn't just ML — it needs behavioural understanding. Adaptability beats static accuracy. Rules + ML + LLM reasoning beat any one alone. Stack: Python, FastAPI, PyOD, NLP, LLM APIs, audio processing, Pandas.
