Deo.Siatah

Year — 2026

Hifadhi app alternative credit scoring

In Kenya, close to 70% of agricultural work happens informally much of it by women, youth, and persons with disabilities. They have the land knowledge, labour, and market relationships. What they do not have is the asset lenders demand: collateral. But these farmers are not starting from zero. They have spent a lifetime building assets the formal system refuses to count: trust, verified local networks, a history of in-kind settlements, savings groups, and a productive farm. Hifadhi 2.0 scores that. We bring together: - Community-verified trust built over a lifetime - Verified local network participation—in-kind settlements and savings groups - Satellite-verified farm activity - Transparent, auditable scoring And we do not hand a farmer a number and walk away. Every score comes with an explanation in their local language, so the farmer understands what drove it and what to do next. What comes next: access through feature phones, and AI voice capabilities in local languages for users who cannot read. For farmers: a route from subsistence to scale. For lenders: a credible way into the informal agricultural economy. For the ecosystem: a trust layer that makes invisible productivity visible.

Live LinkGithub
Hifadhi app alternative credit scoring preview

Stack

- Backend: FastAPI, Python 3.12, SQLAlchemy 2 async, asyncpg, Alembic, Pydantic v2, uv- Database: PostgreSQL on Neon- AI / data: Groq via langchain-groq, Google Earth Engine + Sentinel-2- Frontend: React 19, TypeScript, Vite 8, Tailwind CSS 4, shadcn/Radix- Frontend data/state: TanStack Query, axios, Zod, Zustand, d3-force, Motion, React Router- Testing: Vitest, Testing Library, Playwright, pytest- Deployment: Docker on Render, Vercel, Neon

Key Features

  • - Farmer onboarding and farm registration
  • - Trust-network linking: chiefs, elders, neighbours, buyers, agrovets, chamas, SACCOs, institutions
  • - Verification workflow with participant responses and timeouts
  • - Deterministic credit score with full breakdown
  • - Optional satellite-verified farm activity
  • - Score explanations in English and Kiswahili, with expert and farmer modes
  • - Loan eligibility decision and full loan lifecycle: create, disburse, repay, default
  • - Live trust graph with reduced-motion support and recorded replay
  • - Portfolio and analytics dashboard with fairness indicators
  • - Audit logs, fraud flags, and verifier credibility
  • - English/Kiswahili UI with persisted light/dark theme
  • - Roadmap: feature-phone/USSD access and AI voice narration for low-literacy users

Technical Highlights

  • - Deterministic, versioned credit scores with component-level breakdowns and an audit trail.
  • - Two-axis model: character from community/buyer/discipline/settlement signals, plus optional capacity from satellite farm activity.
  • - Relational trust graph in PostgreSQL—no graph database dependency.
  • - Verification sessions snapshot participants, collect per-participant responses, and handle timeouts.
  • - Verifier credibility, pooled testimony, category-diversity bonus, fraud flags, and least-privilege DB roles.
  • - Bias controls: gender, ethnicity, religion, disability, age, and raw acreage are excluded from scoring.
  • - Privacy-reduced explainability payload sent to a swappable LLM provider.
  • - Earth Engine Sentinel-2 NDVI and cultivation-continuity scoring, optional and isolated behind a provider interface.
  • - Frontend graph state driven by a pure reducer and event log; Zod validation at every API boundary.
  • - Loan lifecycle state machine with explicit conflict handling

My Role

  • Full stack developer