Back to Case Studies
Goal-Based DeFi / AI Agents / Multi-Chain / Non-Custodial2024 – 2025

DeFi for Goals,
Not for Dashboards

A non-custodial DeFi platform where users type a financial goal in plain language and autonomous AI agents execute it across chains. Luvon led product direction, end-to-end design, AI and blockchain architecture consulting, and technical team leadership through launch.

"Most DeFi products ask the user to learn the protocols. DeFiMatrix asks the user to state the goal. The agents do the protocol work."

Role

Product Direction + Design + Architecture

Client

DeFiMatrix

Timeline

2024 – 2025

Visit Live Site

$5M–$10M

$DMX raise across three rounds

4

Entities across SG, LK, UK, USA

3

Rounds at $5M, $7M, $10M valuation

2024–2025

Full product engagement timeline

01 — The Context

DeFi reached a strange ceiling. Total value locked is real and growing, the protocols work, and the math is sound. But the audience that can actually navigate DeFi without help has stayed small for almost a decade. The reason is not technical. It is interface.

A user who wants to earn yield on stablecoins without impermanent loss risk cannot do that by clicking three buttons. They have to learn lending protocols, bridge mechanics, gas, slippage, smart contract risk profiles, and what happens when a peg breaks at 3am. The protocols are open. The path through them is closed.

The Thesis

The next generation of DeFi is not a better dashboard for power users. It is an agent layer that takes a plain-language financial goal and executes it across chains on the user's behalf, without ever taking custody.

Conversational AI

Intent parsing that maps plain language to structured, executable DeFi strategies.

Cross-chain execution

Multi-chain routing across major EVM networks, executed without user selecting a chain.

Non-custodial key management

Scoped, revocable agent permissions. The platform never holds user funds.

02 — The Bet

Three contrarian decisions. Each one held at all three layers.

01

The interface bet

The primary input is natural language, not a form. The user types a goal; the platform routes it to the right agent or combination of agents. That decision changed the entire information architecture. The home screen is a chat. The protocols are implementation details. The user never picks a chain.

02

The trust bet

Goal-based execution by AI agents is a black box by default, and a black box is the fastest way to lose a DeFi user. Luvon designed the platform so every agent action is rendered as a structured, readable explanation — why this protocol, why this chain, what the risk profile looks like.

03

The custody bet

Non-custodial from day one. Agents execute through user-signed permissions with scoped, revocable authority. This is the architectural decision that makes DeFiMatrix a serious DeFi product rather than another centralised app.

03 — What Luvon Owned

Full product engagement

Product, design, brand, and architecture cannot be cleanly separated on a product like this. Luvon was engaged because the four disciplines had to be owned by one team.

Conversational UI

The chat-first goal-capture interface across webapp and mobile. Intent parsing, plan surfacing, and structured approval flows.

Agent Reasoning Surface

Structured explanation cards that render every agent action — what, why, which chains, and what outcome was realised versus projected.

Dynamic Risk Hedging

The live hedging surface that shows current composition, recent adjustments, and the reason for each rebalance in plain language.

Live Opportunity Scanning

The filtered opportunity feed tied to each user's risk tolerance and capital position. Suggestions never auto-execute.

Mastercard Debit Card

Multi-currency accounts, card management, Apple Pay and Google Pay integration, daily limits, and transaction history.

Non-Custodial Key Model

The scoped-permission surface — every agent action generates a user-signed, revocable permission the user can see and cancel at any time.

AI & Blockchain Architecture

Consulted on multi-chain routing logic, agent permission model, and non-custodial key architecture. Not a handoff — design decisions inside engineering decisions.

Technical Team Leadership

Ran the technical team through launch. Product, design, brand, and architecture owned as a single discipline.

"The agent reasoning cards read like a financial advisor's note, not a system log. That distinction is the entire trust position of the product."

04 — The Build

Five surfaces, one product

Each surface answers one question the user actually has. None of them require the user to understand what a chain is.

Chat Interface — Goal Capture

hedge my ETH against a market crash and still earn some yield

Here's a 3-step plan for that goal:

1

Put 40% ETH into a lending protocol for ETH-denominated yield (3.2% APY, low IL)

2

Hedge 40% via a delta-neutral stablecoin position on Arbitrum (4.8% APY)

3

Keep 20% liquid as crash buffer — no lock-in

DESIGN NOTE

Every plan step is a discrete agent task with its own expected outcome, risk profile, and reasoning footprint. No capital moves until the user approves the plan.

Agent Reasoning Card

EXECUTED

Action

Supplied 1.2 ETH to Aave v3 on Arbitrum

Protocol selected

Aave v3 — highest utilisation rate, audit score 9.1/10

Alternatives considered

Compound (lower rate), Morpho (less liquidity)

Expected outcome

3.2% APY · Realised: 3.4% APY (+6bps)

DESIGN NOTE

The card is designed to be skimmed and to reward deeper reading. The first version was too dense. The second was too vague. The shipped version reads like a financial advisor's note, not a system log.

4B

Agent reasoning as a first-class surface

Every agent action lives in a structured explanation card. What the agent did. Why. Which protocols and chains. What the alternative options were and why they were not selected. What the realised outcome is against the projected outcome. The cards are designed to be skimmed and to reward deeper reading.

DESIGN NOTE

The first version was too dense. The second was too vague. The shipped version reads like a financial advisor's note, not a system log. That distinction is the entire trust position of the product.

4C

Dynamic risk hedging

Risk hedging is rendered as a live surface, not a one-time disclosure. As market conditions shift, the agent rebalances within the user's stated risk profile. The interface shows the current hedge composition, the most recent adjustment, the reason for it, and the expected outcome if conditions deteriorate further.

DESIGN NOTE

This surface is the answer to the question every DeFi user asks at 3am: what is the AI actually doing with my money right now.

4D

The Mastercard-tier debit card

On-chain earnings move directly to merchant spend. Multi-currency accounts — USD balance, BTC wallet — surface alongside the card. Daily and withdrawal limits, collateral additions, and transaction history live inside a single card management surface. Apple Pay and Google Pay integration is designed to feel native to a normal mobile banking flow.

DESIGN NOTE

The card is what turns DeFiMatrix from a yield app into a financial operating system.

05 — Tech & Stack

Primary interface

Conversational input — natural language to structured intent

Agent layer

Autonomous agents with goal-scoped, revocable permissions

Custody model

Fully non-custodial — user-signed scoped permissions, no platform custody

Multi-chain execution

Routing across major EVM chains and beyond

Multi-currency accounts

USD · USDT · BTC · ecosystem token balances

Card layer

Mastercard-tier debit · Apple Pay · Google Pay

Corporate structure

Multi-entity across Singapore, Sri Lanka, UK, USA

Token

$DMX — three-round raise at $5M, $7M, $10M valuation

06 — What It Unlocked

DeFiMatrix is the project where Luvon's Web3 specialism and AI agent specialism collapsed into a single design discipline.

Every decision had to hold at three layers simultaneously — the AI behaviour, the on-chain execution, the consumer interface. There is no clean handoff between them. The agent reasoning surface only works if it explains what the agent actually did on-chain. The chat input only works if the intent it captures maps cleanly to a real cross-chain execution route.

Conversational goal-capture pattern

A Luvon reference for any agentic DeFi product. The pattern has been inherited by every subsequent goal-based product in the studio.

Agent reasoning card structure

The design artefact most often cited when other founders ask Luvon how to approach AI legibility inside a consumer product.

Non-custodial scoped-permission surface

Adopted as a default pattern across all subsequent agent work at Luvon. The architecture of trust, not a disclaimer in the footer.

"Most DeFi products ask the user to learn the protocols. DeFiMatrix asks the user to state the goal. The agents do the protocol work."

"Goal-based execution by AI agents is a black box by default, and a black box is the fastest way to lose a DeFi user."

"Non-custodial from day one. The platform never holds user funds. That single architectural decision is what makes DeFiMatrix a serious DeFi product rather than another centralised app."

Start a Project

Building an agentic DeFi product?

We know how the three layers fit together. Let's talk about your product.

Start a Conversation