AI Consulting Accelerator
OngoingAI-assisted consulting workflow that grounds discovery, analysis and delivery in industry standards (IREB, BABOK, BPMN, PMBOK, Scrum) via retrieval-augmented generation. The system produces structured, source-cited drafts from unstructured client input — with 17 named, versioned skills, stateful multi-round engagements, and an Organisational Memory layer that connects internal knowledge to current engagements. Live on AWS since July 2026 as a private single-admin instance (Cognito, no public sign-up).
Technologies
Problem
Business Analysts, Requirements Engineers and project managers spend time on manual framework lookup and building artifacts from scratch. Knowledge from past engagements does not flow automatically into new ones — patterns go unrecognised, and vocabulary gaps prevent business problems from connecting to relevant technical solutions.
Approach
Three-layer workflow (Discovery → Analysis → Delivery) with two memory types: semantic memory (RAG over 23 indexed frameworks, ~5,380 chunks with source citations) and procedural memory (17 named, versioned skills — invoked by name, not by similarity). An Organisational Memory layer connects five sources (skills, projects, technology, competitor and regulatory reports) via a single embedding call per query, with per-source caps and relevance gating that ensures honest "none found" responses.
Result
Live on AWS as a private single-admin instance (Lambda container + API Gateway with Cognito JWT, S3 + CloudFront, own subdomain) and verified end-to-end. 17 skills, 78 pytest tests, 24 golden cases in DE and EN. Covered artifacts: framework Q&A with source citations, business problem structurer, requirements structurer (BR/FR/NFR, INVEST user stories), stakeholder analysis (RACI), roadmap generator, AI readiness, AI governance (EU AI Act / NIST AI RMF), compliance maturity, value articulation, stateful multi-round discovery, engagement lifecycle and report export as Markdown, Word and PDF. An isolated engagement database follows before any real client data.
Learnings
Invoking skills by name rather than by similarity prevents ambiguity and guarantees consistent output contracts. The vocabulary bridge — business-language description metadata in SKILL.md combined with query-side keyword expansion — closes the business-to-technical vocabulary gap without a complex ML pipeline. Per-source retrieval caps prevent dense reports from crowding out actionable skill references. Language drift requires both server-side language detection and explicit output-language instructions in prompts.
Relevance
Demonstrates AI-assisted end-to-end product thinking, RAG architecture grounded in industry standards, clean separation of semantic and procedural memory, and a secured live deployment on AWS — directly relevant for consulting organisations looking to scale BA/RE processes with AI assistance.