PromenautUse case · SDLC intelligence
Requirements & SDLC Intelligence
Most projects fail before code is written: poor requirements, not poor engineering, and a defect fixed in production costs an order of magnitude more than one caught in analysis. Analysts spend weeks reviewing requirements with no systematic quality gate.
From raw input to signed-off output.
$ promenaut run requirements-sdlc
Proven on a live tier-1 bank implementation programme
Parse & analyse
Ingest requirements from any source (XLSX, DOCX, Jira CSV, Confluence, free text), deduplicate across sources, and extract actors, systems, preconditions and acceptance criteria into the knowledge graph. A gap analyser cross-references domain knowledge to surface missing rules, data models and integrations.
Quality & consistency
In parallel: score each requirement across five dimensions (quality, consistency, testability, completeness, traceability), map cross-requirement dependencies, and run exhaustive pairwise comparison to detect duplicates and contradictions.
Decomposition
In parallel: generate Agile user stories with story points, test cases across five categories at P1/P2/P3, and acceptance criteria as BDD Given/When/Then scenarios with a coverage assessment.
Synthesis & sign-off
A traceability checker builds the requirement → story → test → acceptance-criteria chain and flags orphans. A capstone agent weighs all prior signals into a readiness score and a GO / CONDITIONAL / NO-GO call with a risk register and a phased remediation plan. Humans decide; the business remediates and re-runs.
How it actually works.
- Ten specialised agents; the most capable reasoning is reserved only for the final readiness call, where it changes the decision.
- Every finding is cited to a knowledge-graph entry and constrained to graph data, so the analysis does not hallucinate.
- Five-dimension scoring refuses to average away traceability and decomposition gaps: zero acceptance criteria is an automatic fail on testability.
- Runs are versioned with delta tracking, so a re-run after remediation shows exactly what improved.
- Decision-support by design: no automated gate ships anything; the output is handed to people with the stakeholder questions to resolve.
What it produced.
Proven on a live tier-1 bank implementation programme
NO-GO
Readiness call, with the reasons a human can act on
Hidden conflicts
Contradictions caught across documents that manual review had missed
Stories + tests
Generated with a full traceability chain
Minutes
End to end, versus weeks of manual review
Output delivered to senior stakeholders: a readiness verdict, the requirements that block it, and the user stories, test cases and acceptance criteria to move forward.
The parts that break naive tools.
Contradictions hide across documents
A per-document review cannot see that requirement A and requirement F disagree. Exhaustive pairwise semantic comparison caught a production-blocking contradiction the human team missed.
Generic tools see syntax, not domain
A linter finds formatting issues. The domain knowledge graph lets the gap analyser find semantic gaps: a field referenced in acceptance criteria but never defined, an unnamed upstream system.
Artefacts drift out of sync
Stories, tests and criteria are generated independently, so you can get great stories with no tests. The traceability chain surfaces every orphan and broken link.
Synthesis needs judgement, not averaging
The GO / NO-GO call weighs competing signals from eight upstream agents, so it gets the most capable reasoning in the system.
See it on a workflow you own.
Bring a real process. We'll show what your first agentic workflow looks like on the platform.
