PromenautUse case · Schema transformation
Payment Schema Mapping (SWIFT MT ↔ ISO 20022)
Banks hold decades of transformation logic that converts legacy SWIFT MT messages into ISO 20022 XML and internal wrappers, but the mapping rules are largely undocumented and locked in existing systems. This reverse-engineers the field-level rules from real message pairs rather than hand-authoring them.
From raw input to signed-off output.
$ promenaut run payment-schema-mapping
Single MT202 → pacs.009.001.08 run
Schema analysis
A deterministic parser reads the block-structured MT message and the target XSD first, for 100% reliable field extraction regardless of variance, then an enrichment step adds human-readable names, data types and formats. No model is used for the structural parse.
Mapping generation (gate)
Actual source values are correlated against actual target values, the transformation rule for each is inferred, the existing pattern library is loaded as institutional memory, and a calibrated confidence score is assigned. A human approval gate follows before validation begins.
Validation & testing (gate)
Each mapping is replayed against the real output and checked for mandatory coverage, schema compliance, type compatibility, format validation and value reproduction. Twenty test cases and a six-part documentation package are generated. A second gate freezes the mapping version before it is sealed as an immutable snapshot.
How it actually works.
- Confidence bands route the work: 0.95+ auto-approves, below 0.80 requires human review; a non-blocking quality check flags low-confidence clusters.
- A persistent pattern library (date conversion, composite-field split, party resolution, code mapping) is reused across runs: patterns learned on one message type accelerate the next.
- Handles multi-output targets: one source field can map to both an internal wrapper and the ISO 20022 XML.
- Fully database-driven execution with PostgreSQL checkpointing for resume at every stage and live progress over SSE.
- No spec required: the real MT messages and their outputs are the specification.
What it produced.
Single MT202 → pacs.009.001.08 run
100%
Field coverage
97.7%
Average mapping confidence
Perfect
Pass rate on validation checks
~65s
End to end, brief to sealed package
Test cases and reusable transformation patterns generated automatically; the pipeline paused at the human approval gate exactly as designed.
The parts that break naive tools.
Composite fields do not map 1:1
One SWIFT field can pack value-date, currency and amount together; it has to fan out into three separate ISO 20022 fields. A dedicated decomposition pattern handles it.
Reasoning is non-deterministic; parsing must not be
Structural extraction runs on a deterministic parser for 100% reliability; the reasoning step only enriches metadata and correlates values.
Value correlation is the hard cognitive step
Holding the full source and target field sets in mind, tracing values through candidate transformations and calibrating confidence: this one step gets the most capable reasoning in the system.
Trust without a source-of-truth spec
Inferred rules cannot be auto-shipped, so confidence bands, a non-blocking quality check and two hard approval gates keep a human in the loop before anything is sealed.
See it on a workflow you own.
Bring a real process. We'll show what your first agentic workflow looks like on the platform.
