sales-outreach-automation-langgraph

Abstracts CRM connectivity through a base class pattern (src/lead_loaders/base.py) with concrete implementations for HubSpot, Airtable, and Google Sheets, enabling unified lead ingestion regardless of CRM backend. Each adapter implements standardized read/write interfaces that normalize heterogeneous CRM APIs into a common data model, allowing the workflow to operate CRM-agnostically while maintaining provider-specific field mapping and authentication.

Orchestrates the entire lead lifecycle through a LangGraph StateGraph (src/graph.py) that chains discrete processing nodes (src/nodes.py) with conditional branching based on lead qualification scores and data availability. State flows through research → analysis → qualification → outreach generation stages, with each node updating a shared OutReachAutomationState object that persists context across the workflow, enabling resumable and debuggable multi-step automation.

Processes multiple leads sequentially through the workflow with error handling and detailed logging at each step, enabling visibility into which leads succeeded, which failed, and why. The main execution loop (main.py) iterates through leads from the CRM, runs each through the LangGraph workflow, and logs results including processing time, errors, and generated content, providing operational visibility into the automation system.

Collects lead intelligence by scraping LinkedIn profiles, company websites, and social media presence, then aggregates findings into structured research reports. The research node (src/nodes.py) orchestrates multiple external data sources and formats results as context for downstream LLM analysis, enabling personalized outreach based on recent company news, hiring activity, and professional background.

Analyzes enriched lead data using configurable LLM providers (Gemini, OpenAI, Anthropic) to generate qualification scores and detailed analysis reports. The qualification node (src/nodes.py) sends structured prompts (src/prompts.py) containing lead research, company context, and business criteria to the LLM, which returns structured scores (0-100) and reasoning that determines whether the lead advances to outreach generation. Supports multiple LLM backends through a provider abstraction layer (src/utils.py) enabling cost/latency optimization.

Generates customized sales emails, interview scripts, and analysis reports by combining lead research data with structured prompt templates (src/prompts.py) sent to LLMs. The outreach generation node creates multiple content variants (email, call script, LinkedIn message) tailored to the lead's background, company signals, and business context, enabling sales teams to send personalized outreach at scale without manual copywriting.

Exports generated analysis reports and outreach materials to Google Docs and writes qualification results back to the source CRM system. The document generation node creates formatted reports in Google Docs (enabling easy sharing and editing) while the CRM sync node updates lead records with qualification scores, analysis summaries, and generated content, creating a closed loop between automation and sales tools.

Enables non-technical users to customize the entire sales automation workflow by editing business context (company description, value proposition, target criteria) and prompt templates (src/prompts.py) without modifying code. The system reads configuration from environment variables and prompt files, allowing sales operations teams to adjust qualification criteria, outreach messaging, and analysis focus by editing text files rather than Python code.

Maintains workflow state across processing steps through the OutReachAutomationState object (src/state.py) that persists lead data, research results, qualification scores, and generated content throughout the pipeline. This enables debugging individual leads, understanding where processing failed, and potentially resuming workflows after errors, though persistence is in-memory only and requires external storage for true durability.

Abstracts LLM provider selection through a utility layer (src/utils.py) that routes requests to Gemini, OpenAI, or Anthropic based on configuration, enabling cost optimization by using cheaper models for simple tasks (lead research summarization) and advanced models for complex reasoning (qualification analysis). The abstraction handles provider-specific API differences, token counting, and error handling, allowing the workflow to switch providers without code changes.

Organizes LLM prompts into task-specific templates (src/prompts.py) for research summarization, lead qualification, and outreach generation, enabling consistent prompt structure across the workflow and making it easy to adjust prompts without modifying code. Each template includes context injection points for lead data, business criteria, and company information, allowing the same template to be reused for different leads with different data.