SignalScale (Lead Scoring)
A high-throughput predictive machine learning engine designed to score inbound B2B sales leads in real time, featuring calibrated conversion probabilities and 3-tier cohort segmentation.
- Role
- ML Engineer & Full Stack Developer
- When
- 2026 · Production Deployment
- Built with
- PythonScikit-LearnGradient BoostingStreamlitPandasNumPyJoblibRender Cloud
- Links
SignalScale (Lead Scoring)
A high-throughput predictive machine learning engine designed to score inbound B2B sales leads in real time, featuring calibrated conversion probabilities and 3-tier cohort segmentation.
Why it exists
In B2B sales organizations, sales development reps spend up to 60% of their working hours manually triaging incoming leads, contacting unqualified prospects while high-value inbound opportunities grow cold.
Traditional heuristic scoring (e.g. points for opening an email) fails to capture nonlinear feature interactions between company size, website visit dwell time, referral channels, and content engagement.
SignalScale was built to replace brittle manual rule sets with a statistical, calibrated machine learning pipeline that classifies prospects into actionable Hot, Warm, and Cold tiers with sub-15ms inference latency.
What it does
An end-to-end ML scoring dashboard providing real-time inference, batch CSV scoring, and visual probability distributions.
- Calibrated Gradient Boosting Classifier: Trained on historical enterprise lead cohorts using Scikit-Learn's Gradient Boosting classifier with Platt scaling to guarantee that predicted scores reflect true empirical conversion rates.
- Instant 3-Tier Cohort Segmentation: Automatically groups leads into Hot (conversion prob > 70%), Warm (35–70%), and Cold (< 35%), providing immediate sales routing recommendations.
- Interactive Scoring Dashboard: Built with Streamlit, enabling revenue operations teams to tweak lead attributes via interactive sliders and immediately observe the model's prediction.
- Batch Enterprise CSV Upload: Allows sales teams to upload bulk prospect lists (10,000+ rows), transforming and scoring them in seconds with downloadable ranked priority sheets.
What it won’t do
The model and application adhere to strict data privacy and algorithmic transparency boundaries.
- No Uncalibrated Raw Logit Outputs: Raw boosting margins often distort probabilities toward extremes. SignalScale enforces probability calibration so an 80% score represents an actual 8 out of 10 conversion likelihood.
- Zero Persistent Customer PII Storage: All batch CSV processing and slider adjustments occur strictly in transient memory. Uploaded customer records are discarded immediately after inference to prevent enterprise data leaks.
- Rejects Incomplete Core Signals: The preprocessing pipeline rejects records lacking essential behavioral signals rather than silently substituting random averages.
How it works
SignalScale structures lead evaluation through an automated preprocessing and scoring pipeline.
Feature Engineering: The pipeline derives key behavioral interaction signals including Page-Views-Per-Visit, Form-Completion-Velocity, and Lead-Source-Weights from raw telemetry before feeding vectors into the classifier.
Calibration Verification: A reliability diagram (calibration curve) was computed across 5-fold cross-validation, aligning the Brier score to 0.082 to ensure business users can trust score tiers for financial forecasting.
By the numbers
What’s next
- SHAP Value Explainability: Adding local SHAP waterfall plots so sales reps can see the top 3 positive and negative factors driving any individual lead's score.
- HubSpot & Salesforce Webhook Integration: Directly streaming scores into enterprise CRM pipelines via REST endpoints.
Interactive canvas drawing and Convolutional Neural Network inference