AI agent for financial investments

Inspiration

Knowing of issues via colleagues of Risk & Compliance

What it does

It’s a multi-agent “Risk & Compliance Copilot” that ingests onboarding, card, and payment events, detects typology risks (rules, anomaly, graph, ML), maps them to regulatory obligations via RAG, and produces an auditable Risk Card with scores, evidence, and recommended actions. Investigators get explainable alerts with citations to data, policies, and controls, plus a feedback loop that continuously improves rules, prompts, and models.

How we built it

We composed specialized agents—ingestion/PII guard, entity resolution, typology detectors, planner/orchestrator, obligation retrieval, scoring/explainability, and compliance validation—connected by JSON contracts, a vector store for regulations/policies, and a graph + rules + ML stack for signals. A lightweight UI shows Risk Cards and captures investigator labels, which flow into an evaluation set and update prompts/rules via an automated learning loop.

Challenges we ran into

Balancing precision with coverage was hard: deterministic rules avoided misses but spiked false positives until we fused them with calibrated ML, graph motifs, and anomaly scores and enforced evidence citations for every claim. Data quality and privacy constraints required strict schema validation, entity resolution, and PII minimization before any LLM step, plus robust audit trails to keep decisions reproducible under compliance scrutiny.## Accomplishments that we're proud of none