Martimus-built internal system
Hybrid Matching Engine
A hybrid engine: deterministic business rules narrow the field, a language model ranks and explains, and edge cases escalate for review.
Here is the honest version — the problem it was built to solve, what we actually built, the result, how it works under the hood, what it delivers, and where the same pattern fits another business. Nothing is inflated, and anything client-identifying or proprietary is deliberately left out.
The problem
Matching a deal to the right counterparties was slow, manual, and inconsistent.
What we built
A hybrid engine: deterministic business rules narrow the field, a language model ranks and explains, and edge cases escalate for review.
The result
Structured, explainable matching that outperforms a raw model prompt — with humans governing the decision.
How it works
Matching a case to the best of many options is neither a pure-rules problem nor a pure-model problem. Deterministic business rules do what rules do best — hard eligibility, exclusions, and thresholds — narrowing a large field to a valid one. A language model then ranks the survivors and explains each in plain language. Edge cases escalate to a human, and the human governs the final decision. The result is explainable, repeatable matching that a raw model prompt cannot match.
What it delivers
- Deterministic rules for eligibility and exclusions
- Model-based ranking with written rationale
- Human escalation on edge cases
- Explainable, repeatable output
- Governed decisions — never fully automated
Where this fits
Any matching problem where the wrong match is costly and the reasoning has to hold up: a request to a provider, an application to a program, a case to a specialist, a candidate to a role. The engine generalizes well beyond finance.
Capabilities demonstrated
Presented honestly at its true evidence level. Sensitive internals, client-identifying detail, and proprietary logic are deliberately withheld.
FAQ
Hybrid Matching Engine: questions we get asked.
Why not just ask a language model to do the matching?
Because a model alone invents options that do not exist, ignores hard constraints, and cannot show why it chose. The hybrid keeps the constraints in code where they are exact, and uses the model only for the judgment step it is genuinely good at.
Does this only work for lending?
No. The pattern is general: any problem that matches one thing to the best of many — a case to a specialist, an application to a programme, a request to a supplier — has the same shape.