Delphi

Inspiration

Making data-driven decisions is still inaccessible to most people. Building a machine learning model usually requires finding datasets, cleaning data, choosing models, tuning parameters, and interpreting results, all of which take technical expertise and time.

We were inspired by a simple question:

What if anyone could ask a question and instantly get a data-driven answer?

Delphi was built to bridge that gap by turning natural language into a full machine learning workflow, complete with results and explanations. Predicting the future shouldn’t take hours. With Delphi, it takes minutes.

What it does

Delphi turns natural language into a complete, end-to-end machine learning pipeline.

Users can input a simple prompt such as “Predict diabetes risk based on health indicators”, and Delphi will automatically:

• discover relevant datasets
• clean and process the data
• train and compare multiple models
• visualize results and feature importance
• generate plain-English explanations
• provide audio narration of insights
• allow follow-up questions through an AI interface

Instead of requiring technical expertise, Delphi enables anyone to move from a question to a data-driven answer in seconds.

How we built it

We built Delphi as a full-stack system combining frontend UX, backend orchestration, and machine learning.

AI Orchestration: Google Gemini API (function calling agent) Data Processing: Databricks Free Edition, Apache Spark, Delta Lake ML Tracking: MLflow 3 (experiment tracking + model registry) Models: scikit-learn, XGBoost, LightGBM, SHAP, Optuna Voice: ElevenLabs Text-to-Speech API Backend: FastAPI, Python Frontend: React, Vite, Recharts Hosting: DigitalOcean Droplet

Domain: delphi.tech

Challenges we ran into

• Environment setup
  Managing Python versions and ML dependencies (especially with newer versions like Python 3.14) created compatibility issues.

• Frontend-backend coordination
  Synchronizing real-time pipeline updates with UI state required careful architecture and planning.

• Data variability
  Handling datasets with different formats and structures required flexible preprocessing logic.

• UI/UX balance
  Presenting complex machine learning outputs in a simple and intuitive way was challenging but critical.

Accomplishments that we're proud of

• Building a fully functional end-to-end ML pipeline from a single prompt
• Designing a clean, intuitive UI that makes complex systems accessible
• Integrating model comparison, feature importance, and explanations into one seamless experience
• Delivering a system that bridges AI, data science, and user experience

Built With

• databricks
• digitalocean
• fastapi
• gemini
• javascript
• python
• react