SunCulture Transaction Intelligence

SunCulture is a Series-B agtech startup selling solar-powered irrigation to smallholder farmers across East Africa, financed through microloans. To underwrite those loans, the credit team needed a clean view of each farmer's cash flow — but the raw transaction data was a mess: free-text descriptions, inconsistent merchant names, regional abbreviations, and 500+ category targets ranging from "seeds" to "motorbike repair."

The pipeline's job: classify transactions reliably enough that the output could feed a creditworthiness model that decides real loans.

• 7M+ transactions with noisy free-text descriptions and regional abbreviations
• A 500+ category space — pure rules could never cover the tail; pure LLM was too expensive at volume
• Ground truth was scarce: the hand-labeled gold set was small relative to the category space
• A latency budget — new transactions had to score fast enough to inform loan decisions
• A 95%+ accuracy bar for the credit team to trust the signal at all

A three-tier hybrid, each tier handling what it's cheapest and best at:

1. Rule engine (fast path) — exact merchant matches and known prefixes resolve the high-confidence majority at near-zero cost
2. LLM + retrieval (long tail) — for uncertain transactions, retrieve the nearest labeled examples from the corpus and condition the model on them, RAG-style
3. Confidence-gated human review — only predictions below a confidence threshold route to a person

The loop that kept it improving: errors from each batch were hand-labeled and folded back into the retrieval corpus, so the long-tail tier got sharper with every cycle.

A Python REST service classifies individual transactions or batch files:

• Rule layer covers the majority of common transactions with near-zero marginal cost
• LLM + retrieval layer handles novel or ambiguous descriptions
• Confidence gating routes only the genuinely uncertain subset to manual review
• Output integrates directly with the credit-scoring model feeding loan decisioning

The 99% figure comes from a 10,000-item hand-labeled holdout — not training accuracy, not a cherry-picked subset.

• Accuracy was tracked per tier, so the cheap rule path couldn't quietly degrade behind the blended number
• The iterative eval loop (hand-label errors → fold into retrieval corpus) meant the system improved on exactly the cases it got wrong
• Confidence calibration determined the human-review threshold — the −95% review reduction is a direct consequence of the gate, not a separate claim

• 99% accuracy on the holdout, clearing the credit team's 95% trust bar
• 95% reduction in manual review volume
• A cleaner cash-flow signal feeding microloan underwriting
• Faster loan decisions for smallholder farmers across East Africa