Research & Projects

A withdrawal screening platform for virtual asset service providers. Two engines judge each request: a rule engine that traces the fund path upstream and matches laundering patterns such as fan-in collection, rapid fan-out and peeling chains, and a graph neural network (GraphSAGE) that scores the structure around the destination. The model can only raise a release to enhanced review — it never holds or downgrades on its own — and every decision comes with its evidence chain and a draft suspicious transaction report, with a person making the final call. The public demo runs on synthetic scenarios, and the model has not yet been validated on real labelled data. Separately, GCN / GraphSAGE baselines are benchmarked offline on the Elliptic dataset (203k Bitcoin transactions).

A dual-signed credential wallet held by the worker, queried by verification agents that each represent an institution. Every query passes a three-layer policy gate and returns only a boolean and a reason code, never the underlying record. Agents prove whom they act for through GLEIF vLEI credential chains that are re-verified on every query, and an agent never makes the final decision: it produces a recommendation for a named human reviewer whose revocable role credential is sealed into an independently re-verifiable audit trail.

Led team PepeLab in a national competition hosted by the Ministry of Digital Affairs, Taiwan. A patient collecting medication, joining a study or filing a claim usually hands over far more of their record than anyone needs. MedSSI, built on SSI, VC and DID standards, lets the hospital issue one credential that the patient holds in the government's digital wallet; issuance and presentation verification run through the Ministry's wallet sandbox, and a local layer enforces FHIR-path selective disclosure and a minimum identity-assurance level, rejecting any field outside the scenario.

A perpetual CFD engine whose markets reference real-world assets — US equities, gold and crypto — as synthetic exposure: nothing real is issued or held. Opening a position on an RWA market requires an on-chain-verified "qualified investor" credential, equities switch to reduce-only while their market is closed, and an AI agent can trade only inside a session its owner has bounded on-chain. A keeper pushes public market prices into a guarded oracle on a GitHub Actions schedule, and market signals are sold per request over x402 — USDC over HTTP, settled by a public facilitator. A research prototype: testnet only, mock USDC as margin, and self-issued credentials.

Built on the PepeFi perpetual CFD exchange, Mandate Layer bounds what an autonomous agent can do with a user's funds. Before any position opens, the AgentSessionManager contract checks the agent's identity, whether the session was revoked or has expired, the asset allowlist, per-order margin, cumulative budget and leverage, and a session can only close positions it opened itself. The mandate is signed as an EIP-712 credential by an ordinary wallet or an ERC-1271 smart account, margin can be topped up through Base Spend Permissions, and agents pay for market data and signals over x402. It started before the contest as the PepeLab capstone; the submission lists which parts were built during the hackathon.

An ERC-8004 agent identity plus a selectively disclosed mandate: the owner signs the mandate once, and the agent reveals only the fields a counterparty needs, each proven against a Merkle root. On Monad testnet, PassportGate checks the identity, the mandate's status, the Merkle proofs, the agent's signature and a daily limit inside the payment transaction, and a revoked mandate is refused on-chain. The owner's organisation can be verified with GLEIF vLEI credentials against a local KERIA test setup, and any MCP-capable agent can use the passport through four MCP tools, tested with Claude Code.

Developed a decentralized platform for real estate fractional ownership. Implemented PropertyToken for asset digitization and RentalDistributor for automated yield distribution. Integrated an IdentityRegistry to ensure regulatory compliance.

A controlled comparison of quantum and classical classifiers under a 2×2 design: two data sources (a classical non-linearly-separable set, and one generated by sampling a parameterised two-qubit circuit) crossed with two model families. Implemented on Qiskit Aer with QSVM (ZZFeatureMap + fidelity quantum kernel), a variational QNN, and a QSVT-style approximation, against a logistic-regression baseline.