TrackAm

AI-powered financial intelligence for Africans the credit system forgot, from market traders to salaried professionals. The structured transaction history it generates is the missing foundation for fair African credit scoring. Built for the BeOrchid Africa Developers Hackathon 2026 (FinTech track). Solo build by Philip Narteh, Accra.

→ Live app: https://trackam-indol.vercel.app → Backend repo: trackam-api (Spring Boot + Spring AI on Cloud Run) → Architecture: docs/ARCHITECTURE.md

---

Why this matters

85.8% of African employment is informal¹, and Africa processed $1.1 trillion in mobile money in 2024². Most of that activity produces no financial record. The gap reaches beyond the informal economy too. Even formally employed Africans rarely use QuickBooks or Expensify, because those tools assume a financial culture with no MoMo, no mixed currencies, and no side hustle running alongside a salary. Without records, neither the market trader nor the salaried freelancer can prove their financial story to a bank. That missing paper trail is a big part of why credit scoring works so poorly across the continent.

TrackAm starts from where people actually are. Type a transaction in plain language, snap a receipt, or speak it out loud. Over time the app builds the structured, audit-trailed financial history that has always been missing in this region. It begins life as a tracker, and the data it accumulates becomes the foundation for credit that actually works in Africa.

Try it (under 60 seconds)

1. Open https://trackam-indol.vercel.app in any browser.
2. Sign up with email + password (real Supabase auth, your data is yours).
3. On the Add Entry page, type something natural like:

   Bought 3 bags of rice 150 cedis at Makola
   

4. Hit Parse with AI. Watch the violet AI parsed · 0.4s badge appear, then click How I parsed this to see which tokens in your sentence drove which fields.
5. Hit Confirm & Save and the dashboard updates instantly. Now ask the advisor: "Where do I spend the most?"

Cloud Run idles after a quiet period. If the first response feels slow, give it 10–15 seconds to warm up.

If you only read one thing

• A real, deployed MVP. Nuxt 4 frontend on Vercel, Spring Boot 3.4 backend on Google Cloud Run, Supabase Postgres with pgvector. Both tiers are live and real data flows end to end.
• A 4-provider AI fallback chain for text parsing (Gemini Flash-Lite → Groq → Gemini Flash → Cerebras). Groq Llama 4 Scout handles vision for receipts and MoMo screenshots. The advisor stuffs an aggregated, server-scoped summary of your real transactions into the LLM prompt, so answers are grounded in your data without giving the model any way to request someone else's.
• The "AI moment" is visible. Every parse shows a badge with the parse latency and a collapsible reasoning trace highlighting the input tokens that drove each field. No black box.
• 41 backend tests covering JWT validation, the AI fallback chain, prompt-injection guardrails, and the FX currency-match validator that prevents balance corruption.
• Real-world adaptability. Multi-currency support with conversion at the historical rate per transaction. MoMo and cash-first design. Voice input for users who would rather speak than type.

See docs/ARCHITECTURE.md for the full diagram, reliability layer, security model, and end-to-end data flow.

How it works (the "AI moment" flow)

flowchart LR
    classDef step fill:#ecfdf5,stroke:#10b981,color:#064e3b
    classDef ai fill:#f5f3ff,stroke:#8b5cf6,color:#4c1d95
    classDef done fill:#fffbeb,stroke:#f59e0b,color:#78350f

    A["User types:<br/>'Trotro from Madina 5 cedis'"]:::step --> B["POST /api/ai/parse-text<br/>with Bearer JWT"]:::step
    B --> C{{"AiService<br/>callWithFallback"}}:::ai

    subgraph CHAIN["Provider chain · first schema-valid answer wins"]
        direction TB
        P1["1 · Gemini Flash-Lite<br/>cheap + fast"]:::ai --> P2["2 · Groq Llama 4 Scout"]:::ai
        P2 --> P3["3 · Gemini Flash<br/>complex reasoning"]:::ai
        P3 --> P4["4 · Cerebras gpt-oss-120b<br/>different vendor"]:::ai
    end

    C --> CHAIN
    CHAIN --> D["Schema-validated<br/>ParsedTransaction"]:::step
    D --> E["'AI parsed · 0.4s' badge<br/>editable form · reasoning trace<br/>confidence bar"]:::done
    E --> F["Confirm & Save<br/>POST /api/transactions"]:::done

Loading

The confidence bar shows red below 80, yellow from 80 to 89, and green at 90 and above. Every parsed field stays editable before saving, so the user always has the final word.

Stack

• Nuxt 4 · Vue 3 · Tailwind CSS 4 · Nuxt UI 4
• Pinia (state, persisted via pinia-plugin-persistedstate)
• Chart.js + vue-chartjs (income/expense, category doughnut, activity heatmap)
• Supabase JS SDK (auth, with the JWT validated server-side via JWKS)
• PWA-ready via @vite-pwa/nuxt

Local setup

Prerequisites: Node.js 20+, pnpm 10+

1. Copy the env template and fill in your values:

   cp .env.example .env.local

2. Install and run:

   pnpm install
   pnpm dev

App runs at http://localhost:3000 (or 3333 if 3000 is taken). The dev script loads .env.local automatically.

Requires the trackam-api backend (locally on port 8080, or pointed at the Cloud Run URL).

Environment variables

Variable	Description
NUXT_PUBLIC_API_BASE_URL	Backend URL (http://localhost:8080 for dev, Cloud Run URL for prod)
NUXT_PUBLIC_SUPABASE_URL	Supabase project URL
NUXT_PUBLIC_SUPABASE_ANON_KEY	Supabase anon/public key

Deployment

Deploys to Vercel. Push to master and it auto-deploys. Set the three env vars above in Project Settings, pointing NUXT_PUBLIC_API_BASE_URL at your Cloud Run backend URL.

Security headers (X-Frame-Options, X-Content-Type-Options, Referrer-Policy, Permissions-Policy) are applied via vercel.json.

Project structure

app/
├── pages/         # dashboard, add, history, advisor, goals, categories, more, login, onboarding
├── components/    # shared UI
├── stores/        # Pinia: auth, transactions, goals, categories, chat
├── composables/   # useAI, useSupabase, useAuthToken, useVoice, useConfirm
├── utils/         # formatters, forecasting, parseBreakdown (AI moment reasoning)
├── middleware/    # auth.global (route guards, gated on auth._initialized)
└── types/         # TypeScript interfaces
docs/
└── ARCHITECTURE.md  # full architecture + reliability layer + data flow

---

¹ ILO, Women and Men in the Informal Economy, 3rd ed. (2018). 85.8% of African employment is informal. ² GSMA, State of the Industry Report on Mobile Money 2025. Africa processed $1.105 trillion in mobile money in 2024.