Drop in any dataset. Get a statistically-validated insight and a trained model — automatically.
Every team is sitting on data nobody has time to analyze. Hiring a data scientist is slow and expensive. Most “AI for data” tools just describe your spreadsheet or spit out code you still have to run and trust.
Epiphany is different. It actually does the data science.
It’s an autonomous agent that runs continuously in the background and, with no human in the loop, it:
🔎 explores your data → 💡 forms a falsifiable hypothesis → 🧪 proves it with the right statistical test → 🤖 trains a real ML model → 📦 reports the result.
No dashboards to configure. No queries to write. No data scientist to hire. Just upload a CSV — and watch it think.
Epiphany runs a continuous 5-step loop, modeled on how a real data scientist works. Every step is real — not a scripted demo.
| Step | Powered by | What actually happens | |
|---|---|---|---|
| 1 | Trigger | Fivetran | Wakes the moment new data lands. |
| 2 | Explore | Elastic | Discovers your schema and ranks the strongest real signals. |
| 3 | Reason | Gemini + Google ADK | An LLM agent forms one falsifiable, business-relevant hypothesis. |
| 4 | Validate | Python + SciPy | Runs the correct test — χ², t-test, ANOVA, or correlation — on your real rows. |
| 5 | Deploy | GitLab | Trains a real scikit-learn model and opens it as a Merge Request. |
When Gemini and the Google Agent Development Kit are connected, a managed agent dynamically decides which tools to call to answer your question. The result streams to a live dashboard so you can literally watch the agent reason, query, validate, and ship.
🎯 Real statistics, not vibes.
Tests run with SciPy on your actual data. A weak relationship comes back not
significant — it can say “no.” Models are genuinely trained and scored (real
ROC-AUC / R²) and saved as loadable .pkl artifacts.
📂 Works on any dataset, any domain. Upload a CSV and it auto-detects column types and the target, then picks the right test and model — churn, pricing, healthcare, sensors, anything. Zero configuration.
🤖 Genuinely autonomous. A background loop investigates on its own and rotates through questions. The agent chooses its own tools via the Google ADK — agentic, not a fixed script.
🔒 Safe by design. Any agent-generated code is screened by an AST security scanner and executed in a network-isolated, resource-limited sandbox.
🌐 Live and usable. Real sign-in, bring-your-own-data upload, and connect-your-own-API — all from the browser.
| Sign up & sign in | Connect your own providers |
|---|---|
 |
 |
The flow: sign in → your workspace starts empty → upload a CSV (or connect Elastic) → the agent auto-detects the target and starts working → ask it a question in plain English → get a validated finding, a trained model, and the generated code.
▶ Hosted: epiphany-ds.fly.dev — sign up and upload a dataset.
💻 Run locally:
pip install -r requirements.txt
uvicorn app.main:app
# open http://127.0.0.1:8000, sign in, and upload a CSVAdd a free Gemini API key (GEMINI_API_KEY) for
live LLM reasoning, and a Clerk publishable key
(CLERK_PUBLISHABLE_KEY) for real auth — both optional.
Built with Google Gemini · Google Agent Development Kit · FastAPI · scikit-learn · SciPy · Chart.js · Clerk
Released under the MIT License · © 2026 Om Singhal
Data science that runs itself.