Grow with our vaults · Built at Arbitrum London Open House Hack.

(Intentionally there are no technical aspects/terminologies in depth discussed over here, as we all believe code always works but deep down we need more consice understanding of the primitive and our ideation for the product)

To start with, lets checkout an abstract about Banana

The curated vault has become the default way to earn yield on-chain. A depositor hands stablecoins to a vault, a curator decides where that capital goes, and the depositor receives a single liquid share token in return. By early 2026 this pattern had quietly absorbed the majority of net-new deposits in DeFi — Morpho alone held roughly $5.8B in TVL, with risk firms like Gauntlet, Steakhouse, Veda and Sentora running billions in curated strategies on top of it.

But almost all of these vaults share two structural properties that deserve scrutiny. First, they are usually single-protocol: a curator allocates across the markets of one lending platform, not across the whole opportunity set. Second, the curator's power is enforced socially, not cryptographically - the curator is trusted to stay within a risk mandate, and depositors trust the curator's reputation rather than the contract's guarantees.

Banana takes a different position. It is a meta-vault, a single ERC-4626 share that routes USDC across multiple protocols — Aave V3, Morpho Blue, Compound V3, and Pendle fixed-yield PTs — chosen for the best risk-adjusted return at any moment. And it inverts the trust model. The rebalancing decision is delegated to an autonomous off-chain keeper, but every action that keeper can take is bounded by on-chain code it has no authority to change. Allocation caps, an idle-liquidity floor, a per-block churn limit, a NAV circuit breaker, and a strategy-quarantine switch are all enforced inside the vault itself. The agent optimizes; the contract constrains. Depositors don't have to trust the optimizer — they only have to read the limits

DeFi yield aggregation is not new. Yearn invented the category — pool deposits, deploy into yield sources, harvest and compound automatically. The 2024–2026 evolution was the curator model: rather than hard-coded strategies, a human or firm actively manages which markets a vault touches and how capital is sized across them. It is, by most accounts, the single biggest innovation in vaults since Yearn.

The curator model works. It also has two gaps.

First gap: The opportunity set is too narrow. The dominant curated-vault venues are built around a single lending primitive. A Morpho vault allocates across Morpho markets - a Gauntlet vault sizes risk within the universe its host protocol exposes. That is excellent for isolation, but it means a depositor's yield is a function of one protocol's rate environment. When Aave's utilization curve is paying more than Morpho's curated markets, or when Pendle is offering a fixed rate that beats every floating venue, a single-protocol vault simply cannot reach it. The best risk-adjusted yield in DeFi is almost never concentrated in one place.

Second gap: We believe the mandate is a promise, not a guarantee. Curators publish a risk framework and are trusted to honor it. The two leading philosophies here, Gauntlet's simulation-driven position sizing and Steakhouse's conservative asset selection are both off-chain disciplines. They are run by serious teams and have weathered real stress; Steakhouse processed $108.5M in liquidations in a single day in February 2026 while every vault stayed fully redeemable. But the depositor's protection ultimately rests on the curator choosing to behave. The contract does not stop a curator from over-concentrating, draining the liquid buffer, or churning the book. (These are extracted from different research backed sources)

Risk management is far more vital than we think. Any financial service company cant afford a cleanup, some big players hate it even its a small reimbursement, lol. For intance, Quantitative approaches are meaningful and human intervened in most cases. This helps to hand the priority in hands of humans rather than a closed custody benchmarking. We love this approach and are exploring rich tastes around defi risks.

Here at Banana, we love diamond Proxy architecture and standards enhancing code modularity. The contract layer is far more sophisticated in its own way, and to be clear, we are not reinventing the wheel here! We believe in existing development and standards (which has always been a perk being in this industry). But we are struggling with more often market risks evaluation with current standards and innovations being easiliy accessible (most of the services, products are selective in terms of free integration, subscription knots being a bottleneck for engineers to build on top of these protocols). This is expected and we are working to integrate one of these solutions from the research backed institutions.

Some light on the Architecture we aim to scale

We believe every finance/defi instrument should make complete use of the substrate thats being laid down at Banana

The vault is built on the EIP-2535 Diamond pattern. Rather than one monolithic contract, the vault is a single address whose logic is split across facets - independently deployable modules that share one storage space. This matters for a meta-vault specifically: adding a new yield venue is adding a facet, not migrating depositors. A new strategy plugs in through diamondCut without changing the vault address, the share token, or anyone's balance. To make that safe, every module stores its state under an EIP-7201 namespaced slot. Storage collisions — the classic failure mode of upgradeable proxies — are structurally impossible across facets.

Core facets

Vault - the ERC-4626 surface: deposit , withdraw , redeem , share accounting reentrancy guard. This surface is deliberately not casually upgradeable; the asset and share semantics depositors rely on stay fixed while strategies rotate beneath them. And note that these facets are not actually touched by owners or curators but are core part of the underneath base protocol.

a/ AllocatorFacet - the strategy registry and target allocations (in basis points). The owner sets caps and the idle floor; the curator sets targets and rebalances within them.

b/ GuardFacet - the NAV circuit breaker.

c/ FeeFacet - performance fee (capped at 50% of gains, paid only above a high-water mark) and a linear management fee (capped at 10%/yr).

d/ RolesFacet - appoints and revokes curators. The owner is always implicitly a curator, a curator is always strictly less powerful than the owner.

e/ WithdrawQueueFacet / LockFacet - asynchronous exits and an anti-MEV share lock.

Some light on how we revolve around share accounting

Share price is the standard ERC-4626 relationship:

shares = assets * totalSupply / totalAssets

The *totalAssets() sums the vault's idle USDC plus each strategy's live valuation. Each strategy reports its own position honestly: For an instance, Aave and Compound read their rebasing balances directly.

Morpho marks MetaMorpho shares at the current exchange rate; Pendle marks PTs at their oracle-implied discount before maturity and 1:1 after. The vault uses virtual share offsets to neutralize the inflation-attack vector, and fees accrue on every deposit/withdraw against a high-water mark so depositors are never charged twice for the same gain.

We shortlisted three roles, strictly nested to avoid protocol and governance level confusion especially: owner ⊇ curator ⊇ user , and the boundaries are enforced in code, not the convention we laid in ideation or something.

1/ Users deposit, redeem, transfer shares, and queue large exits. Nothing else.

2/ The curator (the keeper, in practice) can only set allocations, rebalance, harvest rewards, fulfill withdrawals, and run the circuit-breaker checkpoint. It cannot exceed a cap, drop below the idle floor, move more than the churn limit in a block, change fees, upgrade a facet, or touch user funds.

3/ The owner configures the limits, registers strategies, manages curators, quarantines venues, and upgrades facets - but cannot directly move depositor capital either.

This is the inversion that defines Banana Markets. The optimizer lives outside the trust boundary. Even a fully compromised curator key cannot rug the vault, it canonly shuffle capital inside a box the owner drew and the code enforces.

Let's also checkout the Offchain base architecture currently holding all strings together and tighter!

The curator is not a desk. It is an autonomous Go agent that wakes frequently and runs one disciplined loop.

a/ Perceive - Read the vault's full on-chain state, then overlay live data from every venue - Aave's liquidityRate and utilization, Compound's Comet rate, Morpho's net APY plus keyless on-chain market risk, Pendle's oracle-implied fixed yield and maturity gate.

b/ Decide - Price each venue's Expected Loss from a closed-form credit model, then water-fill the budget in certain bps increments toward the highest marginal yield i.e λ·EL, never past a cap. Always deterministic, replayable, backtestable. No black box at all.

c/ Validate - Re-check every hard bound off-chain. If a decision would breach one, the agent submits the current targets i.e a deliberate no-op. A bad thought can never become a bad transaction.

d/ Execute - Signing the curator txs before rebalance , harvest , fulfillWithdraw. These are based on independent cadences, with a dry-run mode that decides without sending.

This is a complete water level heuristc , a classic book algorithm. The whole agent (we like to call it dumb agent, since no Large Language Model exists in the architecture yet) has no privileged API and no special channel. It reads what any user reads and sends what any curator could send. There is no off-chain oracle at the Banana Markets as of now (We may add few services as per stratgy fulfillment).

And yes, this is the tornado eye for entire brain of the automated curation that entirely depends on some math that's not over curated.

At a product compounding like this, there is a monolithic need of per asset oracle decentralized service backing, maybe an isolated auction based solver networks and a governance tallying each yeild cycle as a final product. We are interested to keep digging until we find a gem in this space. As we said earlier in above prospect elements, we are working on highly curated risk management modules and tie-ups and we now have an MVP deployed on arbitrum mainnet with all possible ai based audits and formal verification.

~ First Banana intern