Live AI Portfolio Rationalization Advisor Deployed on Streamlit Community Cloud
🚀 View Live App

Executive Summary

The AI Portfolio Rationalization Advisor is a fully deployed Streamlit application demonstrating how a Centaur governance model can be applied to one of the most common and costly problems in enterprise IT: deciding which applications to retain, modernize, retire, or replace.

Enterprise organizations waste significant budget on redundant, aging, and poorly governed application portfolios. The decision process is typically fragmented, spreadsheet-driven, and inconsistent across business units. This application structures that decision into a governed, explainable, auditable workflow, with AI handling pattern recognition and humans retaining full accountability for every outcome.

Built using Python, Streamlit, and the Anthropic Claude API, the app is live at ai-app-advisor.streamlit.app and deployed from a GitHub repository on Streamlit Community Cloud.

What This Demonstrates

Most AI prototypes demonstrate a chatbot. This demonstrates AI product thinking at a Senior PM level: governance, explainability, cost management, and human oversight built into every workflow decision.

Governance by design

Three-layer Centaur model with AI engine, APM review, and CIO escalation. Every decision is logged with reviewer identity, timestamp, and written rationale.

Explainability first

Every AI recommendation exposes the seven signals that drove it, a confidence score, and a Jagged Frontier position showing where AI is reliable versus where human judgment is essential. No black box outputs.

Cost-aware AI

Per-session query limits with a visible quota counter. Access-controlled Screen 6. Deliberate architectural choices that scale without runaway API costs.

Deployment-ready architecture

Live ServiceNow REST API with graceful CSV fallback. Auto-detection at startup. A data mode indicator so users always know what they are looking at.

The Centaur Governance Model

Named after the mythological figure combining human and horse, the Centaur model pairs AI analytical capability with human judgment, accountability, and contextual reasoning. It is the architectural backbone of the application and the answer to the question every enterprise AI buyer asks: who is accountable when the AI is wrong?

🤖
Layer 1 — AI Engine

Analyzes 7 signals per application. Generates a recommendation (Retain / Modernize / Retire / Replace / Evaluate) with a confidence score and a Jagged Frontier position showing where AI is reliable versus where human judgment is essential.

👤
Layer 2 — APM Review

The Application Portfolio Manager reviews every AI recommendation in the Validation Queue. They can approve, override (with mandatory written rationale), or escalate to the CIO.

🏢
Layer 3 — CIO Escalation

High-stakes, high-risk, or ambiguous decisions are escalated to executive review. Every escalation is flagged in the Decision Audit Log with full context.

Why this matters as a moat: Enterprise AI buyers are under regulatory and board-level pressure to demonstrate AI accountability and auditability. The Centaur model delivers this out of the box. Competitors adding AI to existing portfolio tools will struggle to retrofit this. It has to be designed in from the start.

The 6 Screens

Each screen serves a distinct role in the governance workflow. Screens 1–5 are fully public. Screen 6 requires an access key to control API costs.

Screen 1
Portfolio Overview
Executive dashboard

At-Risk Apps, Portfolio TCO, Pending Decisions, and Cycle Accuracy KPIs. Real-time multi-select filters by Lifecycle, AI Recommendation, and Business Unit. Full application table with AI recommendations and status for the current rationalization cycle.

Screen 2
AI Analysis & Signals
Full explainability for every recommendation

Select any application to see the complete signal breakdown: Incident Trend, Duplicate Functionality, Technical Debt Index, Lifecycle Status, 12-Month Incident Volume, Stability Inverse, and Replacement Candidate Exists — each scored 0–100 with a severity badge. Jagged Frontier position (Inside means strong AI reliability, Edge means the APM should validate carefully, Outside means the APM must exercise independent judgment) tells the APM how much to trust the AI output. A plain-English “Why this recommendation?” panel makes the reasoning accessible to non-technical stakeholders.

Screen 3
Validation Queue
The APM's primary workspace

All pending recommendations requiring human review in the current cycle. Three decision actions: Approve, Override (requires mandatory written rationale — a deliberate governance control that forces the APM to articulate their reasoning and creates accountability for future model improvement), or Escalate to CIO. Recommendation Drift Alerts flag applications where the AI recommendation changed since the previous cycle.

Screen 4
Decision Audit Log
The accountability mechanism

A complete record of every governance decision: application name, decision type, AI recommendation at time of review, reviewer identity, timestamp, written rationale, and confidence score. Override rate analysis flags data quality issues or scoring inconsistencies that may require investigation. In production, entries persist to a database and are exportable for compliance reporting.

Screen 5
Governance Dashboard
Portfolio-level analytics for senior stakeholders

Cycle KPIs: total decisions, approval rate, override rate, escalation rate, and cycle accuracy vs. historical benchmarks. Recommendation distribution across Retain / Modernize / Retire / Replace / Evaluate. Frontier Analysis chart showing AI model reliability distribution across the portfolio.

Screen 6
Semantic CMDB Search
AI-powered conversational search — access key required

Natural language queries against the full portfolio dataset, powered by the Claude API via context stuffing — a deliberate architectural choice for explainability at current scale. 20-query session limit with visible quota counter. Password-gated access controls API costs. Eight suggested prompts guide users to high-value queries: cost of all end-of-life apps, dependency risk of retiring a specific application, comparative modernization priority between two systems.

Technical Architecture

Every architectural decision was deliberate, with a documented rationale for when and why it would change at scale.

Stack

  • Python + Streamlit — rapid full-stack deployment without a separate frontend
  • Anthropic Claude API (claude-sonnet) — conversational semantic search on Screen 6
  • ServiceNow REST API — live CMDB data with graceful CSV fallback
  • GitHub + Streamlit Community Cloud — production deployment, CI/CD via git push

Deliberate Choices

  • Context stuffing over vector DB: At 15-application demo scale, stuffing the full portfolio into the Claude context window is faster, more transparent, and easier to debug than a retrieval pipeline. The roadmap documents when pgvector on Postgres becomes the right call.
  • Direct Claude API over LangChain: LangChain abstracts away the reasoning I need to show interviewers. Direct API calls are transparent and controllable.

Data Architecture

  • Five CSV files representing a 15-application enterprise portfolio: application inventory, dependency relationships, security vulnerability records, business capability mappings, 12-month incident summary
  • Auto-detection at startup: live ServiceNow PDI if credentials present, silent CSV fallback if not
  • Data mode indicator in sidebar — users always know their data source

Governance Controls

  • Per-session API query limit (20 queries) with visible countdown
  • Password-gated Screen 6 — session-scoped authentication
  • Mandatory written rationale for every override decision
  • Recommendation change detection across rationalization cycles
  • Full audit log with reviewer identity stamped on every decision

24-Month Product Roadmap

A 0→500 customer scaling strategy with four deliberate phase gates, each unlocked only when the previous phase's key question is answered with paying customers, not assumptions.

Phase 1 · Months 1–4
Validate
3 design partners

User auth, multi-user sessions, decision persistence, prompt A/B testing, Postgres for audit log persistence, CSV import for customer portfolio data. Key question: Will they pay?

Phase 2 · Months 5–10
Launch
10 paying customers

Self-serve onboarding, tiered pricing (Starter / Growth / Enterprise), LangChain for large portfolios, pgvector, model routing — cheap models for simple queries, Claude for complex reasoning. ServiceNow partner ecosystem outreach. Key question: Can we retain them?

Phase 3 · Months 11–18
Scale
50 customers

LangGraph agentic layer (reads and writes to ServiceNow autonomously), drift detection v2, fine-tuned model on anonymized customer data, SOC 2 Type II, SSO/SAML, LangSmith for AI observability. Key question: Can we grow without me?

Phase 4 · Months 19–24
Expand
200–500 customers

Full agentic cycle management, proprietary model trained on anonymized customer data, adjacent markets (cloud cost governance, vendor rationalization, tech debt management), multi-region deployment, Series A or acquisition. Key question: Platform or exit?

The ServiceNow bet: ServiceNow's APM module is expensive and complex. Their mid-market customers, companies with 50 to 500 applications, are underserved. A lightweight, AI-native alternative that integrates with ServiceNow rather than replacing it is a wedge, not a threat. This positions the product for acquisition by ServiceNow or a major SI partner.

Artifacts Produced

🚀
Live Application

Fully deployed on Streamlit Community Cloud. Open in any browser, no install required.

Open App
📄
User Guide

11-page reference covering all 6 screens, the governance model, data sources, and FAQ.

View Guide
📊
24-Month Roadmap

Four-phase 0→500 customer scaling strategy with phase gates, revenue targets, and build vs. buy decisions.

View Roadmap