AI Decision Engine

One decision. 8 AI agents. One definitive report.

---

How Traditional AI Makes Decisions:

• Single-turn chat — guesses an answer from training data
• No data verification — doesn't know if its information is outdated
• Never questions itself — confidently wrong with a straight face
• Gives you a "sounds right" suggestion, but you have no idea what it missed

How the 8-Agent Decision Engine Works:

• 8 specialized AI agents working in concert, each with a distinct role
• Framework first, then research, then modeling, then challenge, then synthesis — every step has a dedicated specialist
• A Devil's Advocate deliberately attacks your top choice to find hidden risks
• A Trend Forecaster looks 5-10 years ahead, not just at today
• Full transparency: you can see every agent's raw output at every step

---

The Decision Pipeline

graph TD
    A["User Input<br/>Describe decision + background + options"] --> B["FRAMEWORK<br/>AI generates analysis framework"]
    B --> C{"User Review<br/>Adjust weights & scenarios"}
    C --> D["ORACLE<br/>Official data research"]
    C --> E["ECHO<br/>Community sentiment"]
    D --> F["MODEL<br/>Quantitative analysis"]
    E --> F
    F --> G["DEVIL<br/>Devil's advocate challenge"]
    F --> H["SAGE<br/>Trend forecasting"]
    G --> I["NEXUS<br/>Final recommendation"]
    H --> I
    I --> J["MIRROR<br/>Process review"]
    J --> K["Final Report"]

    style A fill:#f8fafc,stroke:#e2e8f0
    style B fill:#ede9fe,stroke:#7c3aed,color:#7c3aed
    style C fill:#fef3c7,stroke:#d97706,color:#92400e
    style D fill:#dbeafe,stroke:#2563eb,color:#2563eb
    style E fill:#fce7f3,stroke:#db2777,color:#db2777
    style F fill:#fef3c7,stroke:#d97706,color:#d97706
    style G fill:#fef2f2,stroke:#dc2626,color:#dc2626
    style H fill:#ecfdf5,stroke:#059669,color:#059669
    style I fill:#e0e7ff,stroke:#4338ca,color:#4338ca
    style J fill:#f1f5f9,stroke:#64748b,color:#64748b
    style K fill:#ecfdf5,stroke:#059669,color:#059669

Loading

---

Six Steps, Every One Backed by Evidence

Step 1 — You Talk, AI Listens

Tell the AI what you're deciding. It auto-generates a personalized form based on your decision type — choosing a grad school and choosing a job ask completely different questions.

Step 2 — AI Builds the Framework, You Approve It

AI generates a complete analysis framework: evaluation dimensions, scenario assumptions, weight allocations. Presented in a modal — you adjust the parameters yourself before anything continues. This isn't a black box. You're in the driver's seat.

Step 3 — Dual-Track Research, Cross-Validated

Two agents deploy simultaneously:

• ORACLE checks official data: costs, policies, admission rates, employment stats
• ECHO scans community sentiment: what real people say on forums and social media

Step 4 — Quantitative Modeling, Not Gut Feeling

Based on research data, runs a Monte Carlo simulation:

• Audit layer scoring (each dimension 0-100)
• Multi-scenario analysis (boom / bust / policy shift — all tested)
• Utility function ranking (not just upside — accounts for risk and tail loss)
• Sensitivity analysis (which assumption, if changed, flips the conclusion?)

Step 5 — Challenge + Forecast

Two more agents in parallel:

• DEVIL advocate: attacks the #1 option, defends the last-place option, finds the most fragile assumption
• SAGE forecaster: 5-10 year view on visa policy, industry trends, market cycles

Step 6 — The Verdict

NEXUS integrates all prior agent outputs into a final recommendation. MIRROR reviews the entire process, flags blind spots and information gaps.

Delivers a complete report: conclusion, ranking, rationale, action plan, deadlines.

---

The Agent Team

Agent	Role	What It Does
FRAMEWORK	Framework Architect	Dynamically generates evaluation dimensions, scenarios, and weights based on your situation
ORACLE	Official Researcher	Verifies costs, policies, curriculum, employment data — the hard facts
ECHO	Sentiment Analyst	Collects real reviews from Reddit, forums, social media — what people actually say
MODEL	Quantitative Analyst	Monte Carlo simulation + audit scoring + utility function ranking
DEVIL	Devil's Advocate	Challenges assumptions, finds logical flaws, checks for survivorship bias
SAGE	Trend Forecaster	Analyzes 5-10 year trends in policy, industry, and markets
NEXUS	Decision Architect	Integrates all inputs into the final recommendation
MIRROR	Process Reviewer	Audits analysis quality, flags blind spots and data gaps

---

Quick Start

# 1. Make sure you have Python 3.8+ (standard library only — no pip install)
python --version

# 2. Start the local server
python server.py

# 3. Open in browser
# http://localhost:3456

You need an API key from any supported provider: OpenAI · DeepSeek · Gemini · Anthropic · or any OpenAI-compatible endpoint.

---

Design Principles

Principle	Details
Zero hardcoding	Dimensions, scenarios, and weights are ALL dynamically generated based on your specific situation
Resume from failure	Completed steps are cached — if the pipeline fails midway, it picks up where it left off
Full transparency	Every agent's raw output is visible in the activity panel
Privacy first	Your personal data stays in local input/ (gitignored) — the repo contains only the framework
One-click start	Pure Python standard library, zero dependencies to install

---

Project Structure

index.html          ← Frontend UI (single file, ready to use)
prompts.js          ← 8 agent prompt definitions
server.py           ← Local proxy server (API relay)
agents/             ← Agent role definition docs
templates/          ← YAML templates and examples
input/              ← Your data (gitignored, stays local)
data/               ← Intermediate analysis (gitignored)
output/             ← Final reports (gitignored)

License

MIT