AI Spec-to-Proposal Router

Intelligent model routing engine that analyzes customer specifications, scores complexity, and routes each request to the most cost-effective LLM. Includes a real-time token cost dashboard.

97% cost reduction | 3-tier routing | Real-time dashboard | PDF export

---

What It Does

Instead of sending every request to an expensive flagship model, this system uses a cheap classifier to analyze each request's complexity first, then routes it to the optimal model:

• Simple requests (standard specs) → cheapest model (~$0.0005/request)
• Medium requests (multi-parameter) → mid-tier model (~$0.003/request)
• Complex requests (system integration) → powerful model (~$0.02/request)

Result: up to 97% cost savings compared to a single-model approach.

---

Architecture

flowchart LR
    A[Customer Spec] --> B[Classifier\nGPT-5 Nano]
    B --> C{Complexity?}
    C -->|SIMPLE| D[GPT-5 Nano\n$0.05/1M]
    C -->|MEDIUM| E[GPT-5 Mini\n$0.25/1M]
    C -->|COMPLEX| F[Claude Sonnet 4\n$3.00/1M]
    D --> G[Proposal + Cost Dashboard]
    E --> G
    F --> G

Loading

---

Features

• 3-Tier Model Routing — Complexity-based routing to GPT-5 Nano, GPT-5 Mini, or Claude Sonnet 4
• 16-Product Photonics Catalog — Fuzzy matching across wavelengths, power, applications, and features
• Token Economy Dashboard — Interactive Plotly chart comparing costs across 5 models
• Technical Feasibility Matrix — Green/yellow/red status per requirement parameter
• PDF Export — Professional proposal document with Unicode support
• Webhook / CRM Integration — Simulated webhook trigger for downstream automation
• Privacy Compliance — PII scrubber step before API calls
• Demo Mode — Full functionality without API keys (3 pre-built proposals)

---

Quick Start

# Clone the repository
git clone https://github.com/YOUR_USERNAME/spec-to-proposal-router.git
cd spec-to-proposal-router

# Install dependencies
pip install -r requirements.txt

# Run in demo mode (no API keys needed)
streamlit run app.py

Live Mode (optional)

# Copy and fill in your API keys
cp .env.example .env
# Edit .env with your OPENAI_API_KEY and ANTHROPIC_API_KEY

# Run with live API calls
streamlit run app.py
# Then toggle "Demo Mode" off in the sidebar

---

Project Structure

spec-to-proposal-router/
├── app.py                  # Streamlit main application
├── products.py             # 16-product photonics catalog + search engine
├── pricing.py              # Token pricing models (5 LLMs)
├── agents/
│   ├── classifier.py       # GPT-5 Nano complexity classifier
│   ├── router.py           # Model routing logic
│   └── proposal.py         # Proposal generator + 3 mock proposals
├── utils/
│   ├── cost_calculator.py  # Token cost comparison utilities
│   ├── export.py           # PDF export with fpdf2
│   └── pdf_parser.py       # PDF text extraction (PyMuPDF)
├── styles/
│   └── custom.css          # Custom Streamlit theme (1000+ lines)
├── docs/
│   └── architecture.md     # Detailed architecture documentation
├── .streamlit/
│   └── config.toml         # Streamlit theme configuration
├── .env.example
├── requirements.txt
└── README.md

---

How It Works

1. Complexity Classification

Every incoming customer specification is first analyzed by a cheap classifier (GPT-5 Nano at $0.05/1M tokens). The classifier categorizes the request into three tiers:

Tier	Triggers	Model	Cost/1M tokens
SIMPLE	Standard wavelength, clear power spec	GPT-5 Nano	$0.05 input / $0.40 output
MEDIUM	Multiple parameters, noise specs, multi-wavelength	GPT-5 Mini	$0.25 input / $2.00 output
COMPLEX	System integration, THz, custom solutions	Claude Sonnet 4	$3.00 input / $15.00 output

2. Product Matching

A keyword-based search engine scores each of the 16 products in the catalog against the customer specification. Matching dimensions include:

• Wavelength (exact + range matching)
• Power requirements (mW and W)
• Application keywords
• Feature matching
• Product-specific keywords
• THz frequency matching

3. Token Economy

The dashboard visualizes the cost comparison across all 5 models for the same token count, showing exactly how much the routing approach saves compared to always using the most expensive model.

---

Configuration

Custom Product Catalog

Replace products.py with your own product catalog. Each product is a dictionary with:

{
    "id": "unique-id",
    "name": "Product Name",
    "category": "Product Category",
    "type": "laser",  # or "terahertz"
    "wavelengths": [532, 1064],
    "power_range_mw": [100, 5000],
    "applications": ["Microscopy", "Spectroscopy"],
    "key_features": ["Feature 1", "Feature 2"],
    "keywords": ["keyword1", "keyword2"],
}

Model Pricing

Update pricing.py to match your LLM provider's current pricing.

---

Tech Stack

Component	Technology
Frontend	Streamlit
Classifier	OpenAI GPT-5 Nano
Medium Tier	OpenAI GPT-5 Mini
Complex Tier	Anthropic Claude Sonnet 4
Charts	Plotly
PDF Export	fpdf2
PDF Parsing	PyMuPDF (fitz)

---

License

MIT License — see LICENSE for details.