AI-Powered Lead Qualification Chatbot Using LLM and Embedded Context

Sector: Sales and Marketing
Duration: February 2025 - Present

Work Delivered:

Streamlit Front-End Development:

• Developed an intuitive Streamlit UI where users input natural language prompts.

• Designed clean UX for non-technical users to interact seamlessly with the lead qualification chatbot.

• Enabled fast iteration and testing of user queries without requiring technical knowledge.

Natural Language Preprocessing:

• Implemented a preprocessing layer to clean and normalize user input.

• Applied regex-based sanitization, entity extraction, and noise reduction to improve LLM prompt quality.

• Ensured compatibility with the prompt engineering strategy to maximize LLM performance.

Embedded Database Creation and Management:

• Used Supabase to build and manage an embedded vector database from multi-source inputs: LinkedIn data, CRM exports, and third-party data feeds.

• Transformed raw CRM and LinkedIn data using embedding modules before ingestion.

• Enabled semantic search and similarity-based context retrieval for LLM prompts.

LLM Integration and Contextual Prompting:

• Integrated an LLM with context-aware retrieval to enhance answer quality and relevance.

• Retrieved top-k relevant records based on user intent via vector similarity from the embedded database.

• Generated output classifying whether the lead is qualified, and identified key decision-makers within the organization.

Workflow Automation using n8n:

• Built an event-driven automation using n8n that triggers the LLM workflow based on user prompt submissions.

• Included branching logic for handling multiple input types (e.g., leads vs. companies) and responses (qualified vs. disqualified).

• Connected the workflow to backend APIs for logging and follow-up actions.

Pressure Points / Challenges:

LLM Response Accuracy:

• Early models struggled with understanding vague or ambiguous prompts.

• Added preprocessing and structured context prompts to boost classification precision.

Data Context Relevance:

• Initially returned irrelevant or outdated context from the embedded database.

• Tuned embedding logic and similarity scoring to improve precision in top-k retrieval.

Hallucinated LLM Responses:

• The LLM sometimes generated fabricated company roles or incorrect lead information, especially when context was sparse or mismatched.

• Solved this by:

  • Enhancing embedded database categorization (e.g., tagging records by company, industry, role level).

  • Ensuring top-k retrieval included accurate metadata about leads.

  • Introducing a confidence threshold and fallback logic to ask for clarification when ambiguity remained.

Scalability of Contextual Data:

• Vector database required performance tuning as data volume increased.

• Introduced chunking and partitioning strategies to maintain sub-second retrieval times.

Program/Project Overview:

Scope:

• Automate and scale lead qualification and prioritization using AI.

• Identify qualified leads and decision-makers based on structured and unstructured data inputs.

• Deliver immediate insights to sales and marketing teams for faster engagement.

Collaboration:

• Worked with client’s sales and marketing heads to define qualification criteria.

• Partnered with CRM admins and data providers to standardize input formats for embedding.

Promises:

Automated Qualification:

• All lead prompts analyzed in real-time with LLM-assisted classification.

Standardized Lead Scoring:

• Same logic applied to every user query and lead data, reducing bias and subjectivity.

Faster Time to Engagement:

• Sales reps get near-instant insight into whether a lead is worth pursuing.

Modular and Scalable Design:

• Easy onboarding of new data sources and models with minimal effort.

Problems and Pains (In-Project):

Embedding Data Volume:

• CRM and LinkedIn datasets resulted in large vector sizes.

• Solved via chunking and compression techniques during vector generation.

Prompt Drift:

• Users sometimes input vague or open-ended prompts, which confused the LLM.

• Addressed with a standardized set of instructions in the preprocessing layer.

Hallucinated Responses by LLM:

• Example: The LLM incorrectly stated that a lead was a VP at a company based on unrelated name matches.

• Resolution:

  • Context was strictly scoped using the embedded database's role and organization filters.

  • Only embeddings with verified metadata (e.g., role, company match, email domains) were used to construct prompts.

  • Reduced hallucinations by over 80% after contextual alignment enhancements.

Solution:

• Built a modular architecture where components (frontend, LLM, vector DB, automation) operate independently but integrate seamlessly.

• Used Supabase to manage both the collector database (raw inputs) and the embedded database (vectorized data).

• Streamlined query flow to fetch the most relevant context for LLM-based classification and decision-maker extraction.

Payoffs:

• Improved Conversion Rates by focusing on better-qualified leads.

• Smarter Sales Targeting with key decision-maker insights baked into output.

• LLM Accuracy Confidence: Over 80% reduction in hallucinated or fabricated responses thanks to grounded context.