Contextual Prompt Chaining for Emerging Markets: Inside Gasura's Financial AI

The Challenge: LLMs Fail in Informal Economies

Standard financial budgeting apps assume structured bank feeds (Plaid, Open Banking APIs) and predictable monthly salaries. In Rwanda and across Sub-Saharan Africa, **over 80% of personal transactions occur via Mobile Money (MTN MoMo / Airtel Money)**, communal rotating savings (Tontines / Ibimina), and cash inflows.

When foundation models are asked to budget for informal income, they routinely give western-centric advice (*"Invest in index funds"* or *"Set up auto-debit"*). **Gasura** was engineered to solve this through localized contextual prompt chaining.

1. Multi-Stage Pipeline Architecture

Instead of asking a single prompt to parse, categorize, and advise, we split the workflow into 3 sequential micro-tasks:

terminal

1Raw MoMo SMS / Inflows ──► [ Stage 1: Entity Normalization ]
2                                    │
3                                    ▼
4                          [ Stage 2: Volatility & Buffer Modeling ]
5                                    │
6                                    ▼
7                          [ Stage 3: Localized Advisory Generator ]

1. **Stage 1 — Normalization**: Translates unstructured SMS alerts and informal amounts into normalized Rwandan Francs (RWF) transactions with transaction fee deductions. 2. **Stage 2 — Volatility Engine**: Calculates the household's emergency runway in days rather than months, accounting for seasonal price fluctuations. 3. **Stage 3 — Advisory Formulation**: Employs Gemini 1.5 Flash with strict JSON schemas to formulate practical, micro-action steps (*"Set aside 15,000 RWF for your communal group contribution on Friday"*).

2. Guardrails & Output Validation

We enforce deterministic structured outputs using Google GenAI SDK schema definitions:

typescript

1const responseSchema = {
2  type: "object",
3  properties: {
4    healthScore: { type: "number", minimum: 0, maximum: 100 },
5    bufferDays: { type: "integer" },
6    actionPlan: {
7      type: "array",
8      items: {
9        type: "object",
10        properties: {
11          step: { type: "string" },
12          amountRwf: { type: "number" },
13          urgency: { type: "string", enum: ["high", "medium", "low"] }
14        }
15      }
16    }
17  }
18};

💡 Key Architectural TakeawayImpact: Gasura won First Place at the 2025 Google AI Hackathon in Rwanda for its deep contextual relevance and sub-second advice latency.