iVolve Integration Platform Selection — Agentic Framework

Artefact-driven, citation-backed iPaaS recommendation tool. Ingests SAP PO / SAP NEO CPI integration artefacts, heuristically classifies each interface (business area, EIP pattern, integration style), and produces differential scoring across nine dimensions against seven modern iPaaSes (SAP BTP Integration Suite, Boomi, MuleSoft CloudHub, Workato, webMethods.io, Frends, Azure Logic Apps).

The framework runs as an MCP server driven by Claude Code.

Reliability principle

Every numeric score is computed by deterministic rules over extracted features × KB weights. The LLM is only used to (a) classify ambiguous things and (b) write narrative justification that cites the exact features and KB entries that drove the score. Every output carries an evidence[] array — source artefact path, feature value, KB entry version, citation.

Quick start

# 1. Install
pip install -e ".[dev]"

# 2. Drive it from the web UI (recommended for humans)
.\scripts\webui.ps1                # or:  ipaas-select-webui
# → opens http://localhost:8501 with the Run Wizard

# OR — drive it from Claude Code via the MCP server
ipaas-select-server
# Then in Claude Code call the tools in order:
# start_run → ingest_po_bundle / ingest_cpi_bundle → classify_interface
# → score_interface → score_run_aggregate → render_outputs

The web UI is the friendliest path: it surfaces a 5-page wizard that uploads artefacts, classifies + scores + aggregates, then delivers a research-backed business-case page with interactive heatmaps, per-interface redesign notes, and one-click downloads.

Architecture

See the approved plan for the full design. Layout:

src/ipaas_select/    # Python package
  ingest/            # Deterministic parsers (TPZ / iflow.zip / ESR XML)
  features/          # Deterministic feature extractors
  classify/          # LLM-assisted classifiers (temp=0)
  scoring/           # Hybrid scorer + aggregate + target-platform redesign
  render/            # JSON / XLSX / PPTX / DOCX / PDF renderers
  util/              # Evidence + citation builders
kb/                  # Versioned knowledge base (data, not code)
  platforms/         # One YAML per target iPaaS
  dimensions/        # One YAML per scoring dimension
  patterns/          # EIP catalogue + per-platform suitability
  industry_overlays/ # Scoring sensitisation per industry / OEM affinity
samples/             # Sample TPZ / iflow.zip for tests & demos
tests/               # Unit, snapshot, e2e tests
outputs/             # Per-run output folders (gitignored)

KB hardening cadence

Per the iVolve / Tarento Vision deck, KB content is hardened bi-monthly. Every KB file declares last_reviewed and kb_version. Re-running a scoring run with a newer KB produces a side-by-side diff.