Lead Classification System

Problem#

Inbound business leads were classified through a workflow previously handled manually by multiple team members — a workflow that scales linearly with lead volume and consumes real headcount on a task that's fundamentally a classification problem. This project built a machine learning system to automate that classification directly from real inbound lead data.

My Contributions#

I owned this end-to-end — my first opportunity to build a complete machine learning solution around a real business problem rather than a utility or proof of concept. That included feature engineering on real inbound lead data, model training and experimentation, and designing the evaluation methodology used to validate the system's accuracy.

Architecture#

At a high level: real inbound lead data is transformed into engineered features, fed into a classification model, and evaluated against a methodology designed specifically for this task rather than a generic accuracy check. The system was built as a prototype demonstrating that a meaningful share of what was previously a manual, multi-person classification workflow could be automated.

Technical Decisions#

The project balanced accuracy against practical business requirements throughout — it wasn't built purely to maximize a benchmark number, but to produce something that could plausibly replace real manual effort. That framing shaped the evaluation methodology as much as the modeling work: proving the system was reliable enough to trust with a real workflow, not just accurate on a held-out test set.

Engineering Challenges#

Feature engineering on real, messy inbound lead data was the harder problem — more so than designing the evaluation methodology itself. Working with real business data meant the features available weren't clean or pre-structured for a classification task; extracting useful, reliable signal from it took real iteration, alongside the model experimentation and evaluation work.

Lessons Learned#

Building a high-performing model is only one part of delivering an ML solution. This system achieved approximately 89% classification accuracy and clearly demonstrated that a significant portion of the manual classification workflow could be automated — but it was ultimately never deployed to production, due to deployment constraints and limited resources rather than any shortcoming in the model itself. Deployment, infrastructure, organizational constraints, and available resources can ultimately determine whether a successful prototype reaches production — a lesson that shapes how I scope ML work today: production readiness is a real requirement to plan for from the start, not an afterthought once a model performs well.

Tech Stack#

Feature engineering over real inbound lead data, a classification model, and a dedicated evaluation methodology built specifically for this task. Further implementation detail is not shared here, consistent with this project's proprietary status.

Related Content#

Read more about my professional work at AILACDS, where the scope and outcomes were different — projects that did reach production.