home

How the platform works

three agent suites over a shared core, mirroring the real engagement pipeline — each consumes the previous suite's artefact
DISCOVERY→MODELLING→BUILD

Start here: Architecture & background →

why these agents exist, the target medallion lakehouse on Fabric, and where everything runs

Discovery → live

transcripts, emails, questionnaire → cited requirements catalogue + decision log

Modelling → new

catalogue + source profiling → dimensional design + source-to-target mapping

Build → live

mapping → built, tested Fabric table → pull request

Discovery deep-dive →

the authored walkthrough: stages, agents, guarantees

How the suites connect

Engagement inputs meeting transcripts client emails questionnaire DISCOVERY 6 sub-agents: extract, merge, decisions, contradictions, approvals, gap analysis every item carries a citation MODELLING SourceProfiler — read-only Modeller — dimensional design SourceToTargetMapper profiling is deterministic code BUILD 14-step end-to-end workflow two self-correction loops YAML-compiled DQ tests → tested Fabric table + PR requirements-catalog.yaml + decision log source-to-target mapping files shared core — runtime · base agent · provider-agnostic LLM backends · code-enforced guard rails

Each suite consumes the artefact the previous one produced, the same way a real engagement hands work between phases. Discovery turns meeting transcripts, emails and questionnaires into a citation-backed requirements-catalog.yaml and an ADR-style decision log. Modelling profiles the real source systems and turns the approved catalogue into a conceptual and logical model, then source-to-target mapping files. Build takes a mapping and ships a built, tested Fabric table as a pull request.

Underneath all three sits a shared, provider-agnostic core — the runtime, base agent, LLM client and guard rails. The guarantees are enforced in code, not prompts: no requirement without a verbatim source citation, no decision marked approved without approval evidence, no silent contradiction resolution.

Everything on this console is generated from real artefacts: the engagement catalogues the agents are building and the scored runs from the eval harness.