This is a submission for the GitHub Finish-Up-A-Thon Challenge
What I Built
DiversiFi is an AI-powered stablecoin diversification app built on Celo and Arbitrum. The premise is simple but personal: your stablecoins shouldn't all be pegged to the dollar.
If you live in Kenya — as I do — inflation isn't an abstract macroeconomic concept. It's the gap between what you earned last year and what that money buys today. It's the reason holding savings in a local currency account quietly destroys purchasing power, and why stablecoins feel like a genuine unlock: your savings can actually compound instead of erode.
But even dollar-pegged stables have their own exposure. And if you care about your continent — about African economies developing their own financial infrastructure, about emerging markets building on-chain alternatives to broken legacy rails — then a portfolio that's 100% cUSD is both financially incomplete and ideologically inconsistent.
DiversiFi tries to fix both problems at once. Connect a wallet, pick a financial philosophy, deposit stablecoins into a non-custodial Safe smart account, and let an AI agent rebalance your holdings across regional stablecoins — cUSD (US), cEUR (EU), KESm (Kenya), COPm (Colombia), PHPm (Philippines), cREAL (Brazil) — based on live inflation and economic data.
The agent doesn't just chase yield. It reads governance forums, World Bank inflation feeds, and economic signals to make allocation decisions that reflect both the numbers and the philosophy you've chosen:
- Africapitalism — keep wealth circulating in African economies
- Islamic Finance — Sharia-compliant, no interest-bearing assets
- Buen Vivir — LatAm philosophy balancing material wealth with community wellbeing
- Global Diversification — maximum geographic spread
- Custom — define your own allocation targets This isn't cosmetic. Each philosophy filters which assets the agent can touch, how it weights rebalancing recommendations, and what it rules out entirely. The goal is a tool that reflects how real people in real places actually think about money — not just a generic robo-advisor with a world-map splash screen.
Demo
🔗 Live app: diversifiapp.vercel.app
📦 Repo: github.com/thisyearnofear/diversify
The Comeback Story
DiversiFi started as a hackathon prototype — the kind that works well enough for a 3-minute pitch but quietly falls apart the moment you try to actually use it.
The core flows were broken. The agent could recommend rebalances but couldn't reliably execute them. The permission system — the piece that makes this non-custodial and therefore trustworthy — was wired up but unenforced, which defeated the whole point. The UI showed allocation targets but gave no real-time feedback on what the agent was actually doing. And the financial strategy layer was mostly decorative; it influenced the copy, not the code.
The push to actually finish it came from submitting to the Ethereum México x Bitso Hackathon — a 5-week global build sprint at the intersection of AI, stablecoins, and payments, with Bitso as a key integration partner and 20% of judging weighted on LATAM real-world impact. Having real mentors and a live demo day in front of regulators and fund managers has a way of clarifying what "done" actually means.
Here's what changed:
Execution layer fixed. _executor.ts now correctly bridges the vault service to the chain via Privy smart accounts, with a local dev fallback that doesn't require a full smart account setup to test against.
Permission model enforced. Session signer policies now actually gate what the agent can spend, on which contracts, within what time bounds. The agent cannot exceed user-defined limits. This is the difference between "non-custodial" as a marketing claim and non-custodial as an architectural guarantee.
Strategy wired into agent behaviour. Each financial philosophy now filters and weights rebalance recommendations at the vault.service.ts level. Africapitalism doesn't just change the UI label — it changes which assets the agent will and won't touch.
Real transaction receipts. Transactions now log through OpenClaw with human-readable summaries. Users can see exactly what the agent did, why, and when — not just a tx hash.
Bitso integration. Added Bitso as a payment rail, bridging fiat on-ramps to on-chain stablecoin positions. For LATAM users this matters enormously: getting funds into the protocol shouldn't require already being crypto-native.
Expanded to Arbitrum. Extended beyond Celo to support Arbitrum, broadening the asset universe and giving users access to deeper liquidity pools.
Fee model stabilised. 1% annual management + 10% performance above high-water mark + 0.10% swap spread, now calculated and settled correctly at withdrawal rather than estimated and forgotten.
The project went from a prototype that made a good pitch to something I'd actually trust with a real deposit.
My Experience with GitHub Copilot
I used Copilot Chat throughout the finishing process — primarily for architecture and debugging, less as a code generator and more as a thinking partner when things got tangled.
The most valuable moments were in the permission and execution layers, which are genuinely non-trivial. ERC-4337 smart accounts, session signer policies, Privy's secure enclave model — these interact in ways that aren't obvious, and when something breaks the error messages are often unhelpfully cryptic. Being able to paste a stack trace or policy config into Copilot Chat and get a focused hypothesis about what was failing saved real time that would otherwise have gone into reading SDK internals line by line.
I also used it to pressure-test the security model. Walking through the architecture — user controls Safe, agent signs within policy, no private key on server — and asking Copilot to look for holes surfaced a few edge cases around policy expiry and fallback signing I hadn't thought through carefully enough. Having something push back on your assumptions is underrated.
It's not magic. It didn't know Mento Protocol's quirks or Celo's specific bundler constraints out of the box. But as a tool for reasoning through complex, interlocking systems — rather than just autocompleting boilerplate — Copilot Chat earned its place in this build.
Find me on Farcaster and Lens — always building at the intersection of AI, emerging markets, and on-chain infrastructure.
Top comments (0)