Executive Summary & Desk Takeaways
- 01A typical 24-month multi-account CDR payload contains 2,500 to 6,000 transaction rows—far too large for synchronous LLM inference within a 5-second SLA.
- 02Pure LLM pipelines suffer from non-determinism and 'hallucinated certainty' when merchant strings are truncated or ambiguous.
- 03A hybrid two-stage architecture—deterministic rules and merchant graph for ~80–85% of volume, backed by contextual LLM enrichment for the long tail—delivers sub-5-second latency and reproducible credit audits.
The Seductive Demo vs. Production Reality
Anyone can paste 20 messy bank statement lines—like 'SQ *BLUE BOTTLE SYDNEY' or 'STRIPE* UBER TRIP HELP'—into a frontier language model and get back a remarkably clean table of spending categories. Over the past two years, dozens of fintech prototypes were built on the assumption that credit transaction intelligence is simply a prompt-engineering problem.
In production credit origination, that assumption collapses immediately. When a borrower consents to share 24 months of transaction history across a primary everyday account, a joint offset account, and two credit cards, the payload routinely exceeds 4,000 rows. Sending 4,000 rows through an LLM API takes 30 to 90 seconds, costs dollars per applicant, and worse—can produce slightly different category totals if re-run during a credit audit.
Inside the Two-Stage Hybrid Pipeline
To achieve sub-5-second processing while maintaining 99%+ classification fidelity on Australian CDR feeds, Credit Lenses separates transaction interpretation into two distinct stages.
Stage 1 is a high-throughput deterministic engine backed by an Australian merchant entity registry (covering major supermarkets, utilities, insurers, licensed wagering operators, BNPL clearing houses, and government agencies). Because CDR payloads include structured fields alongside raw narratives, Stage 1 resolves roughly 82% of transaction volume in under 40 milliseconds with zero stochastic variance.
Stage 2 handles the ambiguous 18% long tail: truncated suburban merchants, Pos/Eftpos aggregators (Square, Zeller, Tyro, Stripe), and peer-to-peer reference strings. Here, compact contextual models evaluate not just the text string, but the transaction's amount distribution, time-of-day, and periodicity—Promote-and-Cache governance then writes verified resolutions back into the deterministic lookup layer.
TABLE: Architectural Trade-offs: Pure Rules vs. Pure LLM vs. Hybrid Pipeline
| Dimension | Legacy Regex Rules Only | Pure LLM Prompt Pipeline | Credit Lenses Hybrid Pipeline |
|---|---|---|---|
| P95 Latency (4,000 txns) | 120 ms | 38,000 ms – 75,000 ms | 3,200 ms (< 5s SLA) |
| Long-Tail Merchant Recall | Poor (~71% on SME/Square terminals) | High (~96%) | High (99.2% combined) |
| Reproducibility / Audit | 100% deterministic | Drifts across model temperature/versions | Deterministic cache + pinned model version |
| False-Positive Control | Brittle keyword collisions | Prone to over-confident guessing | Calibrated confidence score + analyst flag |
Why 'Abstention' Is the Most Important Category
When a transaction descriptor is genuinely opaque—for example, a $1,200 transfer labeled simply 'REF 49201'—the worst thing an AI system can do is guess 'Rent' or 'Loan Repayment' with high confidence.
Our pipeline explicitly supports calibrated abstention: surfacing unresolvable material outflows in an 'Unverified Material Commitment' bucket so the credit officer knows exactly which three transactions require borrower clarification.
Industry Sources & Regulatory References
External research, regulatory releases, and market commentary referenced in this briefing:
- Credit Risk Is a Cash-Flow Problem: LLM-Enabled Enrichment in Underwriting — Simon-KucherSource link ↗
- Cash Flow Data and Credit Scoring: Analytics Infrastructure — Celent ResearchSource link ↗