Overview
A no-code platform that profiles any tabular dataset, selects features, trains regression, time-series forecasting or anomaly-detection models and returns results in a single workflow.
0%
Model accuracy
across multiple business datasets
-0%
Time-to-insight
one-click workflow vs manual ML
The problem
Business teams had recurring ML needs — forecasts, anomaly flags, regressions — but every request depended on a data scientist manually profiling data, engineering features and training models. Time-to-insight was measured in days.
Context
Built as a reusable framework that must work on dynamic, previously unseen datasets with different schemas, sizes and quality, and be operable by non-technical users.
Constraints & challenges
- 01Datasets change shape every time — the pipeline cannot assume a schema.
- 02Feature selection and model choice must be automated without silently producing bad models.
- 03Users need results and confidence signals, not hyperparameters.
- 04Forecasting, regression and anomaly detection need different validation strategies.
My role
Designed and built the framework: automated dataset profiling, feature selection, model training and evaluation across regression, time-series forecasting and anomaly detection, plus the one-click workflow for non-technical users.
The solution
Built a dynamic ML pipeline: uploaded data is profiled automatically (types, missingness, cardinality, seasonality), features are selected and encoded, task-appropriate models are trained and validated, and the best candidate is surfaced with plain-language results. The whole flow runs from a single action.
- Automated dataset profiling
- Automatic feature selection and encoding
- Regression, forecasting and anomaly detection
- Task-aware model validation
- One-click workflow execution for non-technical users
System architecture
System architecture
- Source
- Processing
- Model / LLM
- Output
The profiling stage decides which branches run. Each task type has its own validation strategy (e.g. time-based splits for forecasting) so reported accuracy is honest.
Key engineering decisions
- 01
Profile first, then branch
Automated profiling determines the task and preprocessing, which is what makes the framework work on dynamic datasets.
- 02
Task-aware validation
Forecasting uses time-ordered splits; anomaly detection uses unsupervised scoring. One generic cross-validation would overstate performance.
- 03
Plain-language outputs
Non-technical users get results and confidence, not model internals.
Technology
- Python
- Scikit-learn
- XGBoost
- TensorFlow
- Statistics
- Time-series Forecasting
- Anomaly Detection
Results
Achieved 87% accuracy across multiple business datasets and reduced time-to-insight by over 60%, significantly reducing manual ML effort.
Lessons learned
- Automation is only useful if it is honest — validation strategy matters as much as model choice.
- Design for the schema you have not seen yet.
- The best interface for ML is often no interface.
Building something similar?