How I Built Autonomous Under-Collateralized Lending

Overview

I founded Hashstack Finance, a DeFi lending protocol on Starknet L2 that enabled under-collateralized borrowing without credit committees, whitelists, or institutional gatekeeping. The protocol used smart contract constraints to manage risk autonomously — a category where every other player relied on human credit assessment.

Two product innovations — Degen Mode (one-click leveraged execution) and DIAL (a custom interest rate algorithm) — demonstrated original product thinking in a crowded market. I led the protocol from inception through a structured wind-down in 2026. User funds were returned and remain withdrawable; the protocol interface was moved to GitHub Pages so withdrawals stay available with near-zero operational overhead.

1. Industry Context: DeFi Lending in 2020

Decentralized lending is one of the largest sectors in DeFi. Users deposit crypto as collateral, borrow against it, and pay interest — like a bank loan, but executed by smart contracts without a bank.

The dominant protocols — Aave, Compound, MakerDAO — all require over-collateralization. To borrow $75, you must deposit $100. This protects the protocol but creates a capital efficiency problem: you need more money than you want to borrow.

By 2020, this model had reached $10B+ in TVL. It worked, but it locked enormous capital in unproductive collateral positions.

2. The Thesis: Autonomous Under-Collateralized Lending

Several protocols attempted under-collateralized lending before Hashstack. The difference was how risk was managed.

My thesis was that under-collateralized lending could work without human gatekeeping if borrowed capital was constrained to approved deployment channels within the protocol’s integrated ecosystem. The smart contract enforced solvency — no credit committee, no whitelist, no institutional-only access.

This is the same collateral-vs-access trade-off that underlies all lending: how much freedom do you give the borrower while protecting the lender? Hashstack’s answer was maximum leverage, minimum withdrawal.

3. Product Evolution: V1 to V2

The protocol shipped two versions, each representing a different risk calibration.

The V1 model allowed borrowing up to $300 by providing $100 as collateral, with $70 withdrawable and $230 deployed as in-platform trading capital.

The V2 evolution increased leverage from 3× to 5× while removing withdrawal entirely — more capital to deploy, no escape hatch that created risk without user value. Extra supply above the $1,000 minimum reduced leverage automatically (e.g. $2,500 deposit → 2.5× instead of 5×).

The insight: loosening leverage and tightening containment can happen together. The constraint is the product.

4. Degen Mode: One-Click Leveraged Yield

The Problem

Manual leveraged yield in DeFi requires choosing pools, calculating ratios, borrowing, swapping, and providing LP — each step with gas, slippage, and error risk. Most users lack the expertise or patience.

The Product Decision

Degen Mode abstracted that entire sequence into one click: ranked strategies (returns, APR, depth), select, Execute — the protocol handles the rest atomically.

User flow

1. Supply $1,000+ in a single asset
2. Open the Degen tab — strategies displayed with estimated APR and risk
3. Select a strategy, click Execute
4. Protocol auto-borrows $5,000 and deploys in one transaction
5. Track results on Your Borrow

The same abstraction pattern — collapsing multi-step finance into a single outcome — applies well beyond DeFi.

Degen Mode documentation →

5. Interest Rate Design: Build Ambitious, Ship Simple

Aave and Compound use a single floating rate per asset derived from instantaneous utilization. Hashstack needed commitment periods (2 weeks, 1 month, 3 months) with higher APR for longer lockups — a fundamentally different rate model.

DIAL: The R&D Version

I designed DIAL (Dynamic Interest Algorithm for Lending) with term structure pricing, bounded rates, keccak256-based anti-manipulation sampling, and multi-tranche accounting. It explored the full design space of what a commitment-based rate model could look like.

What didn’t survive contact with production: non-deterministic rates hurt integrators, admin-triggered updates were too slow during liquidity squeezes, and multi-tranche randomization raised audit costs beyond what the TVL justified.

The Production Decision

For V1 testnet (July 2023), I evolved the model to a kinked utilization curve — keeping DIAL’s core insights while cutting the complexity.

What survived from DIAL: term premium concept, bounded-rate philosophy, supply/borrow cashflow identity, anti-manipulation intent.

What got cut: pseudo-random sampling, multi-tranche solver, admin-triggered updates, hard rate caps.

Production parameters: base 2% at 0% utilization; 20% at 90%; 100% at full utilization.

6. Protocol Metrics

7. The Platform Bet: Starknet

I migrated Hashstack to Starknet in August 2022 for ZK-proof cost advantages, Cairo’s formal-verification properties for liquidation logic, and ecosystem grant support.

The technical thesis validated — the protocol worked on Starknet, CertiK passed the audit, and utilization ran above industry benchmarks. Cairo’s formal-verification properties were a genuine advantage for liquidation logic. The challenge was Hashstack-specific: a lean team managing a full Cairo migration while simultaneously iterating on product, with limited runway to do both well.

Source: Starknet migration announcement.

8. The Wind-Down Decision

Trigger: Cairo 1.0 deprecation required a full migration or exit.

Decision: Structured wind-down. User funds were returned and remain accessible for withdrawal. The protocol website/interface was moved to GitHub Pages so it stays functional as long as needed with near-zero operational overhead.

Shutting down a protocol you built from zero is a harder product call than launching one. The runway didn’t support both migration and growth, and I wasn’t going to burn capital or user trust pretending otherwise.

9. What I’d Do Differently

Raise more before committing to a major migration. The Cairo migration was the right technical move, but we hit it with a team and runway sized for product iteration, not a full rewrite. I’d secure dedicated migration capital upfront.

Ship the kink model from day one. DIAL was valuable R&D, but it delayed the rate model that actually worked. The lesson: explore the design space in simulation, not in production architecture.

Resource the team for parallel tracks. Hashstack’s product thesis validated — utilization, users, and the autonomous model all worked. The gap was trying to run migration, product iteration, and growth simultaneously on a lean team. With proper resourcing, the outcome could have been different.

Sources

• Hashstack Docs
• Degen Mode
• DIAL (Wayback)
• Production IRM
• DefiLlama
• Starknet migration
• HSTK · Ethereum (Etherscan)
• HSTK · Starknet (Voyager)
• EIP-5299 (GitHub PR)
• Ethereum Magicians
• CertiK report