Inspiration
My extended family was devastated by the 2008 financial crisis. They had invested heavily in a single grain, encouraged by years of strong profits — then an oversupply and weak demand collapsed the price. They burned through savings, pledged their equipment as collateral, and finally liquidated their land and tractors to repay the loans. The debt tore the family apart and ended in divorce years later. Watching people I love lose everything they'd worked a lifetime to build — not from bad farming, but from a price swing no one warned them about — is why I built FinUnity: so a farmer can see the risk before they bet the assets they can't afford to lose.
What it does
FinUnity is an AI‑powered auditing and loan‑feasibility platform for smallholder farmers and the microfinance institutions that serve them. Where a bank sees a farmer with no credit file and says no, FinUnity builds that file from live evidence — food prices, drought/flood indices, satellite crop health, and weather — and turns it into a defensible risk decision. Four specialized AI agents (Market, Land, Satellite, Weather) score a loan independently, then an orchestrator agent negotiates a final verdict, automatically mediating when two agents disagree sharply. The result is a go/no‑go recommendation with a projected ROI and a full stress test — so farmers avoid betting their livelihood on a crop that's about to crash, and lenders can finance the unbanked with their eyes open.
How we built it
FinUnity is a multi‑agent system (built for the Agent Society track) with two front ends over one FastAPI backend: a Streamlit assessment workbench for the field loan officer, and a Next.js 14 + Tailwind institutional portal for the creditor managing the whole portfolio. The four specialist agents each own one evidence stream — Market runs GARCH volatility + Monte Carlo simulation over World Food Programme price history; Land scores drought/flood from PDSI indices; Satellite reads crop health from imagery via Qwen vision‑language models (qwen‑vl) and Roboflow computer vision; Weather pulls live climate risk from Open‑Meteo. An orchestrator collects their scores and, if any two diverge by more than 40 points, runs a mediation pass before issuing the verdict. We run two AI providers by design: Qwen Cloud handles the high‑volume, latency‑sensitive work — news sentiment (qwen‑flash) and satellite vision — while Kimi (Moonshot) handles the deep, long‑context reasoning and writes the human‑readable verdict, and doubles as an automatic fallback if Qwen's quota is exhausted, with no change to the UI. Critically, every assessment ships with a transparent audit trail that reports exactly which data loaded and which agent fell back — so a lender always knows what the model knows and what it doesn't. The platform is deployed live on Alibaba Cloud ECS (Hangzhou region).
Challenges we ran into
I'm a first‑year finance student and a non‑technical solo founder, so I learned software development, ML, and deployment from zero while building this — and finding experienced technical teammates with a limited network was hard. The hardest technical problem was making a multi‑agent system that stays honest when data is missing: rather than hiding gaps, we built the audit trail and the Qwen→Kimi fallback so the system degrades gracefully and tells you it did. Modern AI coding tools (Qwen, Kimi, and others) bridged many of the gaps and let a solo founder ship a deployed, multi‑agent product.
Accomplishments that we're proud of
We're proud that a solo, first‑year founder shipped a deployed, multi‑agent system that matches the original vision: predictive price modeling, satellite vision, Monte Carlo stress testing, and a negotiating agent society — all live on Alibaba Cloud. The audit trail is the piece we're proudest of: it makes the AI's uncertainty visible, which is exactly the trust a real lender needs. The prototype validates projected farming ROI before financing is approved, protecting both borrower and lender.
What we learned
Obstacles became the design. As a finance student with little code, I learned more about AI, ML, and product than I expected — and that transparency beats a perfect score: a model that admits what it doesn't know is more useful than one that pretends. Qwen Cloud was an invaluable build partner, even when not every solution worked first try. Competing alongside experienced teams as a solo founder has been the real prize.
What's next for FinUnity
Beyond the hackathon, we're turning FinUnity into an enterprise decision‑support and auditing tool for NGOs, microfinance institutions, and agricultural lenders — adding real multispectral (Sentinel‑2) NDVI, per‑farm parcel mapping, and a mobile app for field officers — so underserved farming communities can access fair financing without risking the assets they've worked a lifetime to build.
Log in or sign up for Devpost to join the conversation.