Dilly

November 2025. Agentic AI for real-time AML monitoring, onboarding, and compliance. Built at SingHacks — Asia's first fintech-focused agentic AI hackathon, organised by Tenity. 1st place, Julius Bär RegTech Intelligence Challenge.

the problem

Client Lifecycle Management in financial institutions is breaking down under the weight of compliance. The numbers are stark:

• 70% of financial institutions worldwide lost clients in the past year due to slow onboarding — the highest rate ever recorded
• $72.9m per firm is the average annual spend on AML and KYC operations (Singaporean firms: $68.2m)
• $4.6bn in global regulatory penalties in 2024 alone; $1.23bn in the first half of 2025

Compliance teams are buried in manual work — drafting reports, preparing onboarding packs, screening alerts, chasing policy updates across MAS, HKMA, and FINMA. The problem isn't a lack of rules. It's that the rules change constantly and humans can't keep up.

what Dilly does

Dilly is an agentic AI layer for AML compliance and onboarding — built to turn compliance officers from makers into checkers.

Turns makers into checkers — Dilly drafts reports, prepares onboarding packs, and pre-analyses alerted transactions so compliance officers review and approve rather than build from scratch.

Live policy-to-operations bridge — When MAS, HKMA, or FINMA updates a rule, Dilly maps the regulatory change to procedures in real time. No lag between policy and enforcement.

End-to-end support — Onboarding & KYC, screening alerts, periodic reviews, compliance reporting. The full compliance lifecycle in one system.

how it runs

Regulatory changes flow into Dilly's Hybrid Intelligence engine — a combination of rule-based logic and AI assessment — which generates live alerts with human-readable scoring and reporting. A Human-in-the-Loop layer lets compliance officers verify, approve, or override before anything is actioned.

architecture

The system runs on an orchestration engine with three layers:

Rule Engine — 40 rules across three regulatory jurisdictions:

• 15 agents for MAS, 15 for FINMA, 15 for HKMA — each with a dedicated Policy Monitoring Agent

Report Agent — generates audit-ready documentation with policy citations and rationale

Document + Image Agents — 5 document agents (grammar, formatting, indentation, spelling) and 5 image agents (manipulation detection, reverse image search, GenAI image detection) for KYC document verification

the results

Projected impact at Julius Bär scale:

| Metric | Reduction | |---|---| | Time-to-Yes (onboarding TAT) | ▼ 40–60% | | Alert Review TAT | ▼ 50–70% | | False positives | ▼ 30–50% | | Audit prep effort | ▼ 80% |

Annual business impact estimate:

• 150,000 hours saved (~75 FTEs)
• $12–16M OPEX reduction
• Capacity buffer for regional surges without extra hiring
• Lower downside risk via traceable decisions and faster policy adoption

the team

Built with Klaus Zhou (back-end), Richard Xiong (front-end), and Anna Shcherbatiuk (domain expert). I ran product.