Data Analyst to Data Scientist: Mentorship Program
A 13-week structured curriculum designed and delivered to take a working data analyst to data scientist: Python, machine learning, neural networks, executive communication, and a capstone model taken into production.
Engagement overview
A data analyst at a UK-based technology company expressed interest in transitioning to a data science role. I designed and ran a 13-week structured program — weekly sessions, homework, and hands-on projects, culminating in a capstone presentation to company stakeholders.
Sessions ran Tuesday and Thursday each week, with 4–16 hours of total weekly commitment depending on the phase. Throughout, the mentee was asked to communicate learnings to the supervising executive in terms of business application — building the executive communication muscle alongside the technical one from day one.
Company name, mentee name, and supervising executive are genericized. Compensation benchmarks are in GBP.
Curriculum structure
| Month 1 (Weeks 1–4) | Core Data Science Skills — Python fundamentals, statistical analysis, EDA, data cleaning. Week 4 intensive: Pandas, NumPy, Matplotlib. Project: analyze a real company dataset and report findings. |
| Month 2 (Weeks 5–8) | ML & Model Development — ML concepts and algorithms, model building and evaluation, data ethics and responsible AI. Weeks 7–8: deep learning and advanced algorithms, big data technologies. |
| Month 3 (Weeks 9–13) | Practical Application — data science applied to real business problems, data visualization, executive communication, team collaboration. Week 13: capstone project — full data science problem from EDA through stakeholder presentation. |
Capstone model: workforce escalation prediction
The mentee’s capstone was a classification model predicting whether specific client needs would cause service price escalations; with direct operational implications for workforce cost and efficiency.
Initial model (Random Forest): 79% overall accuracy. The mentee correctly identified this as insufficient — not because of the overall accuracy number, but because recall on the minority class (actual escalations) was too low. Recognizing that accuracy alone is the wrong metric when the cost of false negatives is high was a methodologically sound and independently-reached conclusion.
Pivot to neural network: After recognizing the Random Forest’s limitations, the mentee independently decided to try a neural network architecture. This is the right call and they arrived at it themself.
Final model (Neural Network):
| Accuracy | 94.62% |
| Precision | 93.3% |
| Recall | 96.3% |
| F1-Score | 95% |
The mentee also conducted error analysis on the remaining 5% misclassification rate — identifying negative [redacted] values in escalations as the likely source of model uncertainty. This follow-through distinguished them as a data scientist as opposed to someone who stops at the accuracy number.
Deployment steps included containerization with Docker and testing on live data. The model was presented to company stakeholders with planned productization, monitoring strategy, and future enhancement roadmap.
The presentation and model are the mentee’s own work product. They are not reproduced here. Outcomes are included to demonstrate the effectiveness of the curriculum and framework.
Compensation framework
The engagement included a structured compensation plan tied to program completion milestones, with a target earnings range upon successful transition to Data Scientist. Two structures were proposed: milestone-based salary increments at each program phase, or a single increment upon completion. Both included an optional equity component in lieu of partial salary increase.
Teaching materials: model reference guide
The following reference guide was produced as a teaching document for the Month 2 machine learning module. It covers the primary model types a working data scientist should be familiar with, with a summary and use-case rationale for each. Presented as produced — a reference the mentee could return to when selecting models for new problems.
| Model | Summary | Use Case & Selection Rationale |
|---|---|---|
| Linear Regression | Predicts a dependent variable using a linear equation across one or more independent variables. | Understanding relationships between variables and forecasting when the relationship is linear and variables are continuous. Chosen for simplicity and interpretability. |
| Logistic Regression | Binary classification — predicts the probability of a binary outcome based on independent variables. | Binary classification tasks such as spam detection or credit risk scoring. Efficient with a useful probabilistic interpretation. |
| Decision Trees | Tree-like graph of decisions and consequences. Handles both numerical and categorical data. | Classification and regression when interpretability matters and relationships are non-linear. Easy to explain to non-technical stakeholders. |
| Random Forest | Ensemble of multiple decision trees — aggregates results to improve accuracy and control overfitting. | Classification and regression on large, high-dimensional datasets. More robust than individual decision trees with improved accuracy. |
| Gradient Boosting (GBM) | Builds trees sequentially, each correcting errors made by previous trees. | Regression and classification where predictive power is the priority and data complexity is high. |
| Support Vector Machines (SVM) | Finds the optimal hyperplane to separate classes in feature space. | Complex classification tasks on small to medium datasets. Effective in high-dimensional spaces with flexible kernel options. |
| Neural Networks | Layers of interconnected nodes that learn complex patterns through training. | Wide range of tasks from classification to regression. Best for highly complex, non-linear relationships and large data volumes. |
| K-Nearest Neighbors (KNN) | Classifies a sample based on the majority class among its k nearest neighbors. | Classification and regression when no model fitting is needed and data is tightly clustered. Simple and effective for capturing local data structure. |
| MARS | Non-parametric regression using piecewise linear regressions to build flexible models. | Complex, high-dimensional data with non-linear relationships and variable interactions. Adapts to data structure without assuming a fixed form. |
| Principal Component Analysis (PCA) | Dimensionality reduction — transforms data into fewer variables while retaining variance. | Exploratory analysis and preprocessing before applying other models. Reduces noise, simplifies data, identifies patterns. |
| Time Series (ARIMA / Seasonal Decomposition) | Models data indexed in time order; ARIMA uses past values to forecast future ones. | Forecasting financial data, weather, demand, and any data with trends, seasonality, or autocorrelation. |
Each model has its specific applications and is chosen based on the nature of the data, the specific requirements of the analysis, and the desired outcome. The choice often depends on factors such as the complexity of the data, whether the relationship between variables is linear or non-linear, the size and dimensionality of the dataset, and the need for interpretability versus predictive power.
Competency assessment: summary of learning and growth areas
The following was produced as a structured assessment of the mentee’s progress and a roadmap for continued development beyond the engagement. It documents what was covered during the 13 weeks and identifies the areas flagged as priorities for continued growth.
| Areas Covered During Engagement | |
|---|---|
| Business Strategy Integration | Aligning data science projects with business objectives |
| Exploratory Data Analysis | Advanced EDA techniques and expanded analysis methods |
| Model Development | Basic and advanced model iteration, selection, and output analysis |
| Model Testing & Refinement | Error testing, feature testing, assumption checks |
| Deep Learning | Advanced neural network architectures and applications |
| Model Deployment | Prep for deployment, live data testing, spec writing, team collaboration, enhancement planning |
| Multiple Model Development | Creating and productizing second model; planning for third and fourth |
| Big Data Technologies | Exploration of large-scale dataset technologies |
| Dashboard Creation | Building dashboards for existing models; frameworks for future ones |
| Executive Communication | Aligning data science projects with executive-level strategic objectives |
| Identified Areas for Further Development | |
|---|---|
| Advanced Statistical Techniques | Deeper dive into complex statistical methods |
| Data Ethics & Responsible AI | Ethical considerations in data science practice |
| Advanced Visualization | Complex and interactive data visualization tools |
| Version Control (Git) | Team collaboration using version control tools |
| CI/CD for Models | Implementing continuous integration/deployment pipelines for ML models |
| Agile for Data Science | Applying agile project management principles to data science workflows |
| Data Governance & Compliance | Regulatory requirements and best practices in data handling |
| Executive Communication (advanced) | Communicating complex data science to non-technical executives; executive summaries and presentations |
Beyond this engagement
This program is one of several mentorship engagements I’ve run. Others have included: recent graduate to mid-level analyst at a public utility (targeted training on the specific tools and workflows of the role); junior analyst to senior analyst/data engineer; data scientist producing two models taken into production during the engagement; data scientists aspiring to management coached on executive communication and cross-departmental influence.