Lead Intelligence & Territory Routing System

Lead Intelligence & Territory Routing System End-to-end AI pipeline for automated B2B lead enrichment, scoring, and geo-spatial routing BACKGROUND A B2B SaaS company receives approximately 200 inbound leads per day from a web form. Each lead record contains: company name, website URL, country, city, a free-text message, and contact email. Today, a human manually opens each lead, visits the company website, estimates product fit, and assigns the lead to a sales representative. This process consumes 3–4 hours of daily manual effort and introduces routing inconsistencies. You are hired to automate this pipeline entirely. OBJECTIVE Design and build an end-to-end automated pipeline that: ingests raw leads, enriches each record using an LLM, scores leads for sales fit, routes them to the correct representative based on geo-territory rules, writes results to a structured output, and monitors itself with alerts on failure. PIPELINE SPECIFICATION 1 Ingestion Layer Accept leads via webhook or by polling a Google Sheet / CSV. Normalize field names, validate required fields, and deduplicate incoming records before passing them downstream. 2 LLM Enrichment For each lead, use an LLM to: infer the company industry, employee size tier (SMB / Mid-Market / Enterprise), and an ICP fit score from 0–100 based on the company website and free-text message. Extract intent signals from the message. Generate a one-sentence internal summary for the sales representative. 3 Geo-Spatial Routing Map each lead's city/country to a defined sales territory using the provided territory config (JSON). Assign the correct representative. Handle edge cases explicitly: unmapped regions, international leads outside defined territories, and records with missing location data. 4 Scoring & Prioritisation Combine ICP fit score, intent signals, and company size into a normalised final lead score (0–100). Flag any lead with a final score ≥ 75 as high-priority and trigger an immediate Slack notification to the assigned representative. 5 Output & Storage Write fully enriched, scored, and routed leads to a Google Sheet or Airtable base. Each row must contain: all original fields, all enrichment fields, assigned representative, final score, priority flag, and ISO timestamp. 6 Monitoring & Alerting Log each pipeline run: total leads processed, error count, average score, and high-priority count. Send a daily digest summary. If the pipeline encounters a fatal error, dispatch a Slack alert within 5 minutes. SAMPLE DATA PROVIDED TO CANDIDATE 20 mock leads (CSV) – Includes: clean leads, leads with missing city, duplicate entries, leads with vague or ambiguous messages, and one intentional international edge case outside any defined territory. Territory config (JSON) – 5 sales representatives each assigned to a set of countries or regions. Contains intentional gaps — candidate must decide and document how unmapped territories are handled. REQUIRED DELIVERABLES Working pipeline Runnable end-to-end. Must process all 20 sample leads without crashing. Architecture document One-page explanation of design decisions, tool choices, and tradeoffs made. Output sheet Enriched Google Sheet or Airtable base with all 20 leads processed and scored correctly. Error log Document what failed, why, and how it was handled. A blank log is a red flag. EVALUATION CRITERIA LLM prompt design Are prompts structured and output-controlled? Do they handle ambiguous inputs gracefully? Geo-spatial handling Is routing correct? Are missing-city and unknown-country cases documented and handled? Error resilience Does the pipeline fail gracefully? Are retries, fallbacks, or dead-letter queues present? Code / workflow quality Is the implementation readable, modular, and maintainable by another developer? Decision transparency Does the architecture doc explain why choices were made, not just what was built? Output data quality Are scores reasonable? Is rep assignment correct? Are summaries useful to a sales rep? TECHNOLOGY STACK The candidate is free to choose any stack. Typical approaches include n8n, Make, Zapier, Python with LangChain, OpenAI API, Nominatim or Google Maps for geo-lookup, Airtable, Google Sheets API, Slack webhooks, or FastAPI. Stack choice is part of the evaluation — the candidate must justify why they selected their tools.