content: flesh out flagship with the platform's full sub-agent set

Add the data-standardization assistant sub-agent (Rule + ALBERT + RAG hybrid,
90.4% satisfaction), the ~32% cloud-cost reduction from the eval/MLOps layer,
and the Microsoft-workshop persuasion (architect + 7 engineers) to "Why it
matters". Frames the platform as three cooperating sub-agents. IP-safe:
high-level methods and quality metrics only, no code names or weakness numbers.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: kmink3225

_projects/1_ai_platform.md

@@ -15,14 +15,15 @@ mermaid:
 
 **Role:** Technical Lead / Architect  ·  **Stack:** Python, LangChain, LangGraph, Azure OpenAI, Azure AI Search, FastAPI
 
-I architected an enterprise, domain-specific **multi-agent RAG platform** end-to-end, taking it from a single-agent pilot to a company-wide initiative. The platform turns fragmented internal knowledge into a queryable, cited assistant.
+I architected an enterprise, domain-specific **multi-agent RAG platform** end-to-end, taking it from a single-agent pilot to a company-wide initiative. The platform turns fragmented internal knowledge into a queryable, cited assistant, and comprises several cooperating sub-agents — a **knowledge QnA assistant**, a **data-standardization assistant**, and a **code-analysis agent** — over shared Azure infrastructure.
 
 ### Highlights
 
 - **Knowledge QnA chatbot** — a 9 sub-agent **Self-RAG / CRAG** loop with token streaming and source citation. On a 151-query suite it passed all 10 operational metrics: ~98% user satisfaction, **4.66s** average response, 96.9% citation rate, 100% system success. A 50-question, 4-model **LLM-as-judge** evaluation scored 5.0/5.0 on factuality and reasoning.
+- **Data-standardization assistant** — a Rule + ALBERT classifier + RAG hybrid (LangGraph Reflexion loop) that auto-recommends metadata fields. On a 101-query suite it passed all 10 operational metrics: **90.4%** user satisfaction, 3.75s average response, 0% fallback.
 - **Self-built orchestration vs. general-purpose CLI** — benchmarked **up to ~17× lower cost per query** at the top-performing configuration, validated with paired t-test / McNemar / Cohen's d / bootstrap CI over a 6-metric composite.
 - **RAG pipeline** — Parent-Child + contextual chunking, hybrid search (BM25 + vector), child→parent mapping, and reranking to suppress hallucination; a LangChain → LangGraph → Agentic 3-stage orchestration roadmap.
-- **Evaluation & MLOps** — LLM-as-judge auto-scoring (factuality, reasoning, out-of-scope, multi-turn) + architecture A/B benchmarking + metric logging for operations.
+- **Evaluation & MLOps** — LLM-as-judge auto-scoring (factuality, reasoning, out-of-scope, multi-turn) + architecture A/B benchmarking + metric logging for operations, cutting estimated cloud operating cost by **~32%**.
 
 ### Architecture
 
@@ -58,4 +59,4 @@ Orchestration follows a deliberate LangChain → LangGraph → Agentic roadmap,
 
 ### Why it matters
 
-A self-built harness keeps the control plane in-house: vendor flexibility, cost control, and knowledge captured as a durable asset rather than rented from a single provider.
+A self-built harness keeps the control plane in-house: vendor flexibility, cost control, and knowledge captured as a durable asset rather than rented from a single provider. I made this case beyond my own team, too: across two Microsoft workshops I led the technical discussion and persuaded a Microsoft architect and seven engineers of the self-built orchestration approach over a general-purpose Copilot CLI.