Demo

Test.html page

Prototype UI for testing

Outbound Email

Overview

This project is a prototype AI Agent Stack for detecting business activity signals on SMB websites. It focuses on three high‑value triggers:

• Expansion → “Grand Opening”, “New Location”
• Hiring → careers/job postings
• Scheduler → Calendly / Acuity / “Book Appointment”

Each signal collected by the Scraper Agent is wrapped into an Evidence Card evaluated by the validator agent , then passed to an Outbound Agent that drafts a personalized email.(only if the score passes above the fixed threshold value)

Goal. Help sales and marketing teams discover timely triggers that indicate a business is actively growing and worth contacting now. Please refer flow.md for codeflow and schemas.

Features

• Multi‑agent pipeline with LangGraph orchestration
• Modular design with Pydantic schemas
• Local inference with Ollama
• Scoring heuristics
  • Regex‑based intent detection
  • Confidence weighting with freshness decay
• Automatic email draft generation tied to detected signals
• Configurable via YAML (configs/config.yml)

Tech Stack

• Agents: LangGraph (StateGraph orchestration)
• LLM Runtime: Ollama (local models like phi3.5, llama3.1) #This prototype uses llama3.1 1B param model pulled using ollama.
• Schemas: Pydantic
• Scraping: Requests + BeautifulSoup4 (via tools/)
• Validation: Regex patterns + freshness‑decay scoring

Setup

1. Clone repo

git clone https://github.com/<your-username>/<your-repo>.git
cd <your-repo>

2. Install dependencies

python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

3. Install & run Ollama

• Download Ollama
• Pull a model (e.g., llama3.2:1b) you can plug and play with the models.

ollama pull llama3.2:1b

Ensure server is running locally at http://localhost:11434.

4. Configure YAML

Update configs/config.yml as needed:

gate:
  min_confidence: 0.6

confidence:
  weekly_decay: 0.85
  floor: 0.3

llm:
  validator:
    model_id: llama3.2:1b
    backend: ollama
    max_new_tokens: 200
    temperature: 0.05
  outbound:
    model_id: llama3.2:1b
    backend: ollama
    max_new_tokens: 300
    temperature: 0.25

5. Run pipeline

python -m src.app --csv data/test_sites.csv --vertical #vertical_name --out data/results.jsonl (or results.json)

Testing

Create a test HTML page with signals:

Example:
<h1>Grand Opening! Now Open our new clinic in Cary</h1>
<p>Book an appointment today!</p>
<p>We are hiring staff for our new team.</p>

Serve it locally:

python -m http.server 8000

Add to data/test_sites.csv:

domain,company
localhost:8000,Test Business

Run the pipeline again—expect a populated Evidence Card and email draft.

Output Example

{
  "domain": "localhost:8000",
  "company": "Test Business",
  "card": {
    "signal_type": "expansion",
    "canonical_url": "http://localhost:8000/",
    "snippet": "Grand Opening! Now Open our new clinic in Dallas",
    "confidence": 0.72,
    "explain": "explicit_phrase; freshness=0.92"
  },
  "email": {
    "subject": "Congrats on your new Dallas clinic",
    "body": "Hi Test Business — noticed your Dallas opening.\n\nMany teams use AI-assisted workflows to route leads in real time and convert faster.\n\nWould a 10-minute walkthrough next week be useful?",
    "call_to_action": "Open to a quick walkthrough?"
  }
}

Architecture Notes

• LangGraph coordinates a StateGraph of tools/agents that:
  1. Fetch and parse site content,
  2. Detect signals via regex/semantic cues,
  3. Score with freshness decay,
  4. Assemble Evidence Cards,
  5. Draft outbound emails conditioned on the detected signal(s).
• Pydantic enforces strict schemas for Evidence Cards and email payloads.
• Ollama enables local inference for validator and outbound generation, minimizing cost and maximizing privacy.

Future Improvements

• (ongoing) Create a Natural language understanding interface that maps user prompts to pre-defined ICPs across common verticals.
• Semantic scoring with FAISS + reranker (still rethinking on how to go about it)
• CRM/marketing integrations for automated workflows.

Appendix

Directory Layout

.
├── configs/
│   └── config.yml
├── data/
│   ├── test_sites.csv
│   └── results.jsonl
├── index.html
├── requirements.txt
├── src/
│   ├── app.py
│   ├── agents/
│   ├── tools/
│   └── schemas/
└── README.Rmd

CLI Usage

python -m src.app \\
  --csv data/test_sites.csv \\
  --vertical #vertical \\
  --out data/results.jsonl