How can an LLM assist with unstructured lead interpretation while keeping qualification logic deterministic and external communication under human control?
Why This Problem Matters
Business automation often mixes two different functions: interpreting unstructured input and making binding decisions. Allowing a language model to perform both makes the system harder to audit and can create unnecessary risk in customer-facing workflows.
Approach
The workflow uses an LLM for structured interpretation of a submitted enquiry. Deterministic rules then calculate a lead score and select the High Value or Normal branch. Both branches save structured information and generate a Gmail draft, but a human decides whether the message is edited, approved, sent or discarded.
The core control pattern is:
AI interprets → deterministic logic decides → automation prepares → human authorizes.
Architecture
The n8n workflow connects lead intake, AI extraction, deterministic scoring, conditional routing, Google Sheets persistence, branch-specific response generation and Gmail draft creation.
Evaluation
The public repository documents validation of both Normal and High Value lead paths through representative scenarios and screenshots. The validation demonstrates the MVP workflow behavior; it is not a production reliability benchmark.
My Contribution
The repository is presented as Yasir Siddiq’s engineering portfolio project and contains the sanitized public workflow and supporting documentation.
Limitations
Current limitations include static scoring weights, no duplicate-lead detection, Google Sheets rather than a full CRM, imperfect AI extraction risk, no multi-step follow-up lifecycle and no production-grade retry/error handling.
Reproducibility
The repository provides the sanitized n8n workflow JSON and setup guidance. Users must reconnect their own OpenAI and Google credentials; no production credentials should be committed to the public repository.
Current Status
Functional MVP with validated Normal and High Value routing, structured extraction, deterministic scoring, Google Sheets logging, Gmail draft generation and human-reviewed outbound communication.