8-agent AI architecture accelerator that transforms enterprise discovery inputs into reference architectures, ADRs, governance controls, a 30/60/90 implementation roadmap, and a TCO/ROI/NPV business case.
Live demo: https://danvzla.github.io/ai-architecture-advisor/ · Part of the AI portfolio by Daniel Mazzini
Every enterprise AI engagement starts the same way and takes too long:
| Step | Deliverable | Who produces it |
|---|---|---|
| Day 1–2 | Discovery & problem framing | Senior Architect |
| Day 2–3 | Reference architecture design | Principal Architect |
| Day 3–4 | Architecture Decision Records (ADRs) | Principal Architect |
| Day 4–5 | Risk register & governance controls | Risk Architect |
| Day 5–6 | 30/60/90 day delivery roadmap | Senior TPM |
| Day 6–7 | ROI model & CFO executive narrative | Value Engineer |
5–7 days. Each engagement starts from scratch. Discovery context gets lost in handoffs. ADRs are almost never written until an audit forces it. The ROI case gets assembled after the architecture decision — making it advocacy, not analysis.
This tool solves all three problems simultaneously.
Seven enterprise-grade deliverables from a single 7-step discovery session:
| # | Deliverable | What it contains |
|---|---|---|
| 1 | Reference Architecture | Layered target-state diagram with specific technologies per layer |
| 2 | Architecture Decision Records (ADRs) | Context · decision · rationale · tradeoffs · alternatives rejected |
| 3 | Risk Register & Governance Controls | Risks ranked by severity, category, likelihood, mitigation, and owner |
| 4 | 30/60/90 Day Delivery Roadmap | Phased milestones, deliverables, dependencies, go-live criteria |
| 5 | TCO / ROI / NPV Model | Full financial model with sensitivity analysis at 65%, 100%, 125% |
| 6 | CFO Executive Narrative | Investment justification, urgency driver, recommended decision |
| 7 | AI Readiness Scorecard | 6-dimension maturity score with gap analysis and legend |
Each agent is a separate file with a single responsibility: one prompt, one output schema, one parsing function. Agents run sequentially and each one receives the accumulated outputs of all prior agents as context.
Agent 01 — Discovery Agent
Problem framing · stakeholder mapping · pain quantification
Output: problemStatement · rootCauses · stakeholderMap · urgencyRating
Agent 02 — Outcome Agent
Business outcomes · KPI definition · success criteria
Output: outcomeHierarchy · successCriteria · adoptionMilestones
Agent 03 — Capability Agent
AI capability mapping · pattern selection · UX design
Output: recommendedPattern · capabilityMap · agentDesign · uxRecommendation
Agent 04 — Architecture Agent ← receives prior[discovery] + prior[capability]
Target-state architecture · ADRs · integration patterns
Output: targetArchitectureLayers · adrs · integrationPatterns
Agent 05 — Data & Platform Agent ← receives prior[architecture]
RAG pipeline design · data readiness · platform architecture
Output: ragPipelineDesign · platformArchitecture · integrationDesign
Agent 06 — Governance Agent ← receives prior[architecture] + prior[data]
Risk register · governance controls · compliance framework
Output: riskRegister · governanceControls · complianceRequirements
Agent 07 — Roadmap Agent ← receives prior[governance].riskRegister
30/60/90 day roadmap · milestones · team structure
Output: roadmap · goLiveCriteria · teamStructure · criticalPath
Agent 08 — Value Engineering Agent ← receives prior[ALL] + pre-calculated financials
CFO narrative · investment justification · advisory path
Output: executiveSummary · cfoNarrative · investmentJustification · advisoryPath
The financial model (TCO, ROI, NPV, payback, cost of inaction) is calculated by financial.js — a pure deterministic function — before Agent 08 runs. Agent 08 receives those numbers and writes the CFO narrative around them, ensuring the math is always accurate.
The full blueprint loads in ~8 seconds using a pre-built Telco NOC scenario embedded in assets/data/demo-blueprint.js. All 8 agents animate in sequence. All 7 tabs populate with real, structured content. Works offline, works from the desktop, works anywhere.
Select Demo → choose a scenario → click Quick Demo or Start Wizard.
Live generation using Anthropic's claude-haiku-4-5. Each of the 8 agents makes an independent API call and results chain forward. Live structured generation grounded in your discovery inputs.
Switch to Claude → enter your sk-ant-... key → run.
Identical 8-agent pipeline running on gpt-4o-mini. Same output structure, same 7 tabs. Provider is swapped by one line in api-client.js.
Switch to OpenAI → enter your sk-... key → run.
Each scenario is a complete set of 40+ wizard answers covering a specific industry and use case. Select one and the wizard auto-fills — or fill it manually for a custom engagement.
| # | Scenario | Vertical |
|---|---|---|
| 01 | Telco Network AIOps Platform | Telecommunications |
| 02 | Financial Services GenAI Assistant | Financial Services |
| 03 | Healthcare Clinical Decision Support | Healthcare |
| 04 | Government Document Intelligence | Public Sector |
| 05 | Enterprise IT Operations (ServiceNow) | Technology |
| 06 | Retail Customer Experience AI | Retail |
| 07 | Manufacturing Predictive Maintenance | Manufacturing |
| 08 | Cybersecurity Threat Intelligence | Technology |
| 09 | Insurance Claims Processing AI | Financial Services |
| 10 | Energy Grid Optimization AI | Energy |
| 11 | Legal Contract Intelligence | Professional Services |
| 12 | Supply Chain Intelligence | Manufacturing |
| 13 | HR Talent Intelligence Platform | Technology |
| 14 | Pharma Clinical Trials AI | Healthcare |
| 15 | Private Cloud AI Infrastructure | Technology |
The wizard mirrors how a senior architect actually runs a discovery session:
| Step | Focus | Key inputs |
|---|---|---|
| 1 — Profile | Context | Industry · function · company size · AI maturity (1–6) · urgency |
| 2 — Problems | Discovery | Problems to solve · pain level · stakeholders · problem description |
| 3 — Outcomes | Value | Target outcomes · success metrics · impact timeframe · adoption target |
| 4 — Capabilities | Design | Required AI capabilities · autonomy level · UX · multi-agent need |
| 5 — Data | Platform | Knowledge sources · data quality · deployment model · integrations |
| 6 — Governance | Risk | Industry risk · governance maturity · auditability · internal capability |
| 7 — Value | Economics | Annual volume · effort · labor cost · error rate · implementation cost |
Step 7 inputs drive the financial model: labor savings, rework reduction, delay avoidance, TCO, ROI, NPV, payback period, and cost of inaction.
┌─────────────────────────────────────────────────────────────────┐
│ PRESENTATION LAYER │
│ index.html — mode selector, scenario dropdown, wizard shell │
│ assets/css/styles.css — full design system, dark theme │
│ Wizard overlay · Agent pipeline animation · 7-tab results │
│ Browser-native · GitHub Pages · zero infrastructure cost │
└──────────────────────────┬──────────────────────────────────────┘
│
┌──────────────────────────▼──────────────────────────────────────┐
│ ORCHESTRATION LAYER │
│ assets/js/app.js — mode management, pipeline loop, export │
│ AGENT_REGISTRY[] — ordered agent array (10 lines) │
│ Prior results accumulated and passed forward to each ag
D789
ent │
│ FinancialModel.calculate() runs before Agent 08 │
└──────────────────────────┬──────────────────────────────────────┘
│
┌──────────────────────────▼──────────────────────────────────────┐
│ AGENT + API LAYER │
│ agent-01-discovery.js → agent-08-value.js (8 individual files) │
│ assets/js/api-client.js — Claude ↔ OpenAI provider swap │
│ assets/data/demo-blueprint.js — pre-built output, no API key │
│ assets/data/scoring.js + financial.js — pure math functions │
└─────────────────────────────────────────────────────────────────┘
Each agent file exposes the same interface:
window.Agent04Architecture = {
id: 'architecture',
name: 'Architecture Agent',
mission: 'Target architecture · ADRs · integration patterns',
tokens: 4500,
buildPrompt(answers, prior) {
// Assembles prompt from wizard answers + all prior agent outputs
// Agent 04 receives prior[discovery] + prior[capability]
},
async run(apiKey, answers, prior) {
const raw = await window.ApiClient.call(apiKey, this.buildPrompt(answers, prior), this.tokens);
return window.ApiClient.parseJSON(raw);
},
};Prior results chaining — each agent sees everything prior agents produced:
- Agent 07 (Roadmap) reads
prior.governance.riskRegisterto build phase-specific risk notes - Agent 08 (Value) reads
prior[ALL]plus pre-calculated financials to write the CFO narrative
Provider abstraction — one line change in app.js switches the entire pipeline:
window.ApiClient.init('claude'); // or 'openai'JSON repair chain — 5 attempts to handle malformed or token-truncated responses:
- Direct
JSON.parse() - Strip non-printable control characters
- Collapse literal newlines inside strings
- Collapse whitespace runs
- Close unclosed brackets (handles token-limit truncation)
ai-architecture-advisor/
├── index.html ← deployed file (built, self-contained)
├── build.js ← build script for local development
│
└── assets/
├── css/
│ └── styles.css ← full design system
│
├── data/
│ ├── agents.js ← agent registry (10 lines)
│ ├── scenarios.js ← 15 complete scenario answer sets
│ ├── demo-blueprint.js ← pre-built Telco NOC blueprint
│ ├── scoring.js ← AI readiness scoring engine
│ └── financial.js ← TCO / ROI / NPV calculator
│
├── agents/
│ ├── agent-01-discovery.js ← problem framing prompt + parse
│ ├── agent-02-outcome.js ← KPI definition prompt + parse
│ ├── agent-03-capability.js ← capability mapping prompt + parse
│ ├── agent-04-architecture.js ← architecture + ADRs prompt + parse
│ ├── agent-05-data-platform.js ← RAG pipeline prompt + parse
│ ├── agent-06-governance.js ← risk register prompt + parse
│ ├── agent-07-roadmap.js ← 90-day roadmap prompt + parse
│ └── agent-08-value.js ← CFO narrative prompt + parse
│
└── js/
├── api-client.js ← Claude ↔ OpenAI provider abstraction
├── wizard.js ← step navigation + field rendering
├── pipeline.js ← agent pipeline animation
├── results.js ← all 7 tab renderers
└── app.js ← init · orchestrator · export
Maintenance rule: To add a 9th agent — create agent-09-xxx.js and add one line to agents.js. No other file changes needed.
Running the Telco Network AIOps scenario (Scenario 01) produces:
Financial model
- Annual gross benefit: $845K (labor + rework + delay savings)
- 3-year ROI: 87%
- NPV at 8% discount rate: $1.1M
- Payback period: 14 months
- Cost of inaction (3 years): $2.6M
Architecture pattern selected
RAG + Agentic AI with Human-in-the-Loop and Zero Trust Enterprise Integration
Sample ADR produced (ADR-002)
Decision: Primary: Azure OpenAI gpt-4o. Fallback: Anthropic Claude claude-haiku-4-5 via Azure Marketplace. Rationale: Azure OpenAI meets data residency compliance. Dual-provider fallback ensures 99.9% AI availability with seamless failover. Tradeoffs accepted: Prompt compatibility testing required across both providers. Alternatives rejected: OpenAI direct API (data leaves Azure boundary). Single provider (unacceptable availability risk for 24/7 NOC).
- Node.js (any recent version)
- A text editor
# 1. Clone or download the repo
git clone https://github.com/danvzla/ai-architecture-advisor.git
cd ai-architecture-advisor
# 2. Edit any source file in assets/
# - Add or modify an agent prompt: assets/agents/agent-04-architecture.js
# - Update a scenario: assets/data/scenarios.js
# - Change styling: assets/css/styles.css
# - Modify the financial model: assets/data/financial.js
# 3. Build the self-contained deployment file
node build.js
# → produces dist/index.html (self-contained, works everywhere)
# 4. Test locally
# Open dist/index.html in any browser — no server needed
# 5. Deploy
# Upload dist/index.html to GitHub as index.htmlOpen assets/js/app.js and change one line:
window.ApiClient.init('claude'); // Anthropic claude-haiku-4-5
window.ApiClient.init('openai'); // OpenAI gpt-4o-miniRun node build.js after any change.
| Layer | Technology |
|---|---|
| UI | Vanilla HTML · CSS · JavaScript (no framework) |
| Icons | Tabler Icons |
| LLM (Claude mode) | Anthropic claude-haiku-4-5 (claude-haiku-4-5-20251001) |
| LLM (OpenAI mode) | OpenAI gpt-4o-mini |
| Hosting | GitHub Pages (static, zero cost) |
| Build | Node.js build script (no bundler dependencies) |
The application separates discovery, agent orchestration, deterministic calculations, structured outputs, and human review. Eight agents run sequentially, while financial calculations are completed by deterministic JavaScript functions before the value narrative is generated. The application does not approve investments, architectures, controls, or implementation plans on behalf of an organization.
Deterministic: TCO, ROI, NPV, payback, cost-of-inaction calculations, readiness scoring, scenario structure, pipeline order, required fields, and export formatting.
AI-generated: problem framing, architecture recommendations, ADR drafts, governance narratives, roadmap explanations, and executive summaries.
Current controls include required-field validation, structured agent schemas, sequential context chaining, deterministic financial calculations, JSON parsing and repair, and human inspection of final outputs.
Security note: Do not enter production or long-lived API credentials into the public browser demonstration. Demo Mode requires no API key.
Do not submit confidential customer, employee, financial, security, or regulated data. Production deployment requires a secure backend, managed secrets, authentication, authorization, encryption, audit logging, and data-retention controls.
Current testing is primarily functional and manual: wizard behavior, scenario loading, pipeline sequencing, calculations, result rendering, export, provider switching, and responsive behavior. Production use requires automated unit, schema, prompt-regression, adversarial, accessibility, performance, and security testing.
The application creates structured first-draft decision-support artifacts. Recommendations depend on user inputs and may be incomplete or inaccurate. Demo scenarios and financial figures are illustrative. Enterprise identity, persistent storage, approval workflows, observability, and production integrations are not included.
This project is provided for demonstration and educational purposes and does not constitute architecture, financial, legal, compliance, security, procurement, or investment advice.
Daniel Mazzini — Principal Solutions Architect & Senior Technical Program Manager
20+ years across cloud infrastructure, networking, security, and telecommunications. Background includes Broadcom/VMware, Dell EMC, Juniper Networks, and Nokia/Ericsson engagements with AT&T, T-Mobile, and Verizon.
- LinkedIn: linkedin.com/in/daniel-mazzini-22059734
- GitHub: github.com/danvzla
- Location: Coppell, TX · Bilingual: English & Spanish