⚙️ MLOps · Sequence Modelling · Production AI · GCP Vertex AI
MLOps Sales Pipeline
Intelligence System
Replaced subjective sales rep estimates with a production-grade Hidden Markov Model on GCP Vertex AI, delivering 80% forecast error reduction across 100,000+ leads with 12-month advance signals.
Hidden Markov Model
GCP Vertex AI
Salesforce
Apache Airflow
MLOps
Docker
Python
BigQuery
Production Deployed · Live on GCP Vertex AI
End-to-end MLOps pipeline operational · Airflow DAGs orchestrating retraining · Scoring 100,000+ leads in production
80%
Forecast Error Reduction
5x
More Accurate Than Sales Reps
12mo
Advance Signals Generated
100K+
Leads Scored in Production
86%
Probability Accuracy · Strict Validation
🎯 Business Problem
The sales organisation was consistently over-forecasting deal closure. Sales reps estimated close rates of 41–44% against actual performance of 30–32%, a 10 percentage point gap that compounded into poor quota setting, misallocated headcount, and unreliable revenue projections at the executive level.
The underlying issue was that pipeline reviews relied entirely on rep judgment, ignoring the behavioural signal already sitting in Salesforce stage progression data. Every deal leaves a sequence of stage transitions that tells a more accurate story than a gut-feel estimate. This project captured that signal and operationalised it into a production forecasting system.
🏗️ System Architecture
☁️
Salesforce
Stage Progression
Deal History
→
🐍
Data Pipeline
Python · Pandas
Sequence Filter
→
🧠
HMM Model
hmmlearn
Stage Transitions
→
🐳
Docker
Containerised
CLI Pipeline
→
⚡
Airflow DAGs
Orchestration
Retraining
→
☁️
Vertex AI
GCP Production
Scoring Engine
Validation Layer: Dual regime testing: Strict (6,000 clean leads) · Full population (100,000+ leads) · FY23–FY25 backtesting
⚙️ Key Technical Challenges & Solutions
CHALLENGE
Non-linear sales cycles: deals moved erratically between stages, making raw sequence data unreliable for modelling.
SOLUTION
Training data curated by filtering for clear, directional progression patterns. HMM trained on meaningful transitions, not noise.
CHALLENGE
Variable deal velocity: different time-in-stage durations across opportunity types distorted transition probabilities.
SOLUTION
Dual validation regimes: strict testing on 6,000 clean leads and full population testing on 100,000+ to report generalisation honestly.
CHALLENGE
Production operationalisation: moving from a research model to a retrainable, maintainable system running on GCP at scale.
SOLUTION
Modular CLI pipeline with train/evaluate/score commands, containerised in Docker, orchestrated via Airflow DAGs on Vertex AI.
CHALLENGE
Proving model superiority: demonstrating consistent outperformance over experienced sales reps across multiple fiscal years.
SOLUTION
Multi-year backtesting across FY23–FY25 comparing model vs. rep vs. actual. 5x error reduction confirmed across all periods.
📊 Model Performance vs Sales Reps
| Fiscal Year | Model Forecast | Reps Forecast | Actual | Model Error | Rep Error |
|---|
| FY23–Present | 33% | 41% | 31% | 2pp | 10pp |
| FY24 | 35% | 46% | 44% | 9pp | 2pp |
| FY25 | 34% | 42% | 36% | 2pp | 6pp |
Strict Validation
86% probability accuracy on 6,000 clean, directional leads
Full Population
77% accuracy across all 100,000+ opportunities
Error Reduction
10pp → 2pp average, 80% reduction vs. rep baseline
💡 What Makes This Different
Sequence-Aware Modelling with HMM
- Models temporal stage transition sequences, not static deal attributes
- Captures behavioural momentum that point-in-time models miss entirely
- Each deal scored based on its own unique progression pattern
12-Month Advance Signals
- Closure probabilities generated up to 12 months ahead
- Enables proactive quota setting and resource allocation
- Removes end-of-quarter scrambling from the sales process
Dual Validation Regime
- Strict validation on 6,000 clean leads: 86% accuracy
- Full population on 100,000+ leads: 77% accuracy
- Both regimes reported transparently, no cherry-picked benchmarks
Full MLOps Production Stack
- CLI-driven train/evaluate/score pipeline, not a notebook
- Dockerised and orchestrated via Airflow DAGs on Vertex AI
- Built for continuous retraining at enterprise lead volumes
🛠️ Technical Stack
🧠Hidden Markov Model
hmmlearn · Sequence Modelling
☁️GCP Vertex AI
Production Hosting · Model Serving
⚡Apache Airflow
DAG Orchestration · Retraining
🐳Docker
Containerisation · CLI Pipeline
☁️Salesforce
Data Source · Stage Progression
🐍Python · Pandas · NumPy
Data Processing · Feature Engineering
📊Matplotlib · Seaborn · Plotly
Visualisation · Performance Reporting
🗄️BigQuery
Data Warehouse · Lead Storage
📈 Business Impact
📉 80% Forecast Error Reduction
- Rep error reduced from 10pp to 2pp on average, a 5x improvement
- Directly improves revenue planning and board-level forecast confidence
- Consistent accuracy maintained across FY23, FY24, and FY25
⚡ 12-Month Early Decision Making
- Resource allocation and quota decisions made 12 months earlier
- Hiring plans aligned to pipeline signals, not last-minute estimates
- Moves the sales organisation from reactive to strategically planned
🏢 Enterprise-Scale Production System
- Scoring 100,000+ leads continuously in production
- Retrains automatically via Airflow on new Salesforce data
- Built to scale with GCP infrastructure as lead volumes grow
⚠️ Disclaimer: This case study is sanitised for confidentiality. No client identifiers, schemas, or proprietary visuals are disclosed. Metrics are reported at aggregate levels only.