Skip to content

README.md

Latest commit

 

History

History
161 lines (118 loc) · 5.73 KB

README.md

File metadata and controls

161 lines (118 loc) · 5.73 KB
Raw

Project SPEAR — Strategic Personalized Engagement & Acquisition Resource

An autonomous AI engine that replaces a 5-person marketing team — generating hyper-personalized B2B outreach emails at 10x the volume and 20% of the cost.

Problem

EXL Services identified 5,000+ high-value targets across major banks — but the manual team could only reach 500 leads/month (10% of TAM). To hit volume, they relied on generic templates yielding <1% response rates.

Result: 90% of leads left untouched. The other 10% burned with bad outreach.

Solution

Project SPEAR automates the entire outreach pipeline with 5 AI agents, each mirroring a real marketing team role:

Agent Role Tool
The Research Lead Scrapes LinkedIn via Google, extracts structured lead profiles Serper API + GPT-4.1-mini
The Knowledge Specialist Retrieves the most relevant EXL case studies for each lead ChromaDB + text-embedding-ada-002
The Account Strategist Crafts a seniority-aware strategy brief (pain point, value angle, tone) GPT-4.1
The Writer Drafts a 120-170 word personalized email with strict anti-hallucination rules GPT-4.1
The QA Lead Scores on a 10-point rubric; loops back to the Writer if score < 7 (max 3 tries) GPT-4.1 + LangGraph conditional routing

Pipeline Architecture

graph TD;
    A[config.json — Company, Roles, Country] --> B[Web Scrapper]
    B -->|leads_final.csv| C[ChromaDB Retrieval]
    C --> D[Strategy Agent]
    D --> E[Email Drafter]
    E --> F[LLM Judge]
    F -->|Score >= 7| G[Approved — Save to CSV]
    F -->|Score < 7| E
Loading

Orchestrator: src/main.py runs all stages sequentially via subprocess. Agentic loop: Built with LangGraph — the Judge node conditionally routes back to the Drafter until quality threshold is met or max iterations hit.

Example Output

Input: Chintan Singh, VP Quantitative Research @ JP Morgan

Strategy Brief:

Pain point: Managing model performance and risk at scale while keeping trading algorithms competitive. Lead with: EXL's Quantitative Trading Strategies work — mirrors his world of back-testing and automation. Tone: Peer-to-peer, analytical. He's a quant — no fluff.

Generated Email:

Subject: Automating quant signal generation — what we built for a hedge fund

Hi Chintan, Moving from manual back-testing to fully automated signal generation is one of the bigger operational challenges for quant research teams running multiple strategies simultaneously...

Judge Score: 7.0/10 — Approved after 2 iterations.

Impact

Metric Manual Team (Current) Project SPEAR (Future)
Monthly Outreach 500 Leads 5,000 Leads
Personalization Low (Templates) High (1-to-1 Strategy)
Monthly Meetings ~5 ~50
Annual Pipeline $2.5M per Cohort $25M per Cohort
Cost $30,000/year $5,000/year

Even at a conservative 20% close rate, Project SPEAR has the potential to generate $5M in net new revenue annually.

Tech Stack

  • Orchestration: LangGraph (stateful agent graph with conditional routing)
  • LLMs: Azure OpenAI GPT-4.1 (strategy, drafting, judging) + GPT-4.1-mini (data extraction)
  • Vector DB: ChromaDB with cosine similarity search
  • Embeddings: Azure text-embedding-ada-002
  • Web Scraping: Serper API (Google search for LinkedIn profiles)
  • Language: Python, Pandas

Project Structure

Project_spear/
├── src/
│   ├── main.py               # Orchestrator — runs the full pipeline
│   ├── scraper.py            # Stage 1: Scrape LinkedIn leads via Serper
│   ├── enricher.py           # Stage 2: Enrich leads (Apify + Serper backfill)
│   ├── database.py           # Stage 3: Embed case studies into ChromaDB
│   └── agent.py              # Stage 4: LangGraph agentic email generation
├── data/
│   └── case_studies/         # 9 EXL case study markdown files
├── samples/                  # Example pipeline outputs (see below)
│   ├── leads_raw_sample.csv
│   ├── leads_final_sample.csv
│   └── email_results_sample.csv
├── assets/                   # README images
├── docs/                     # Presentation deck
├── config.json               # Run parameters (company, roles, models)
├── requirements.txt          # Python dependencies
├── chroma_db/                # Generated vector database (gitignored)
├── outputs/                  # Generated CSVs (gitignored)
└── .env                      # API keys (gitignored)

Setup

pip install -r requirements.txt

Create a .env file with your API keys:

AZURE_OPENAI_KEY=your_key
AZURE_EMBEDDING_KEY=your_key
SERPER_API_KEY=your_key

Edit config.json to set your target company, roles, and country, then run:

python src/main.py

Sample Outputs

The samples/ directory contains example pipeline outputs so you can see what the system produces without running it:

File Description
samples/leads_raw_sample.csv Raw scraped leads with name, title, department, and LinkedIn URL
samples/leads_final_sample.csv Enriched leads with headline, skills, and backfilled fields
samples/email_results_sample.csv Generated emails with judge scores, feedback, and iteration counts

Author

Justin Varghese — Data Scientist LinkedIn | GitHub