Guide5 min read

Agentic Data Science — Quickstart

Get your first autonomous data-science project running in 5 minutes. The Agentic Data Scientist (ADS) plans, executes, self-heals, and produces a cryptographically-sealed evidence bundle without manual orchestration.

Prerequisites

A RadMah AI account (sign up at app.radmah.ai/signup)
25 credits — free tier comes with this out of the box
A CSV or Parquet dataset under 100 MB

Upload a dataset

Dashboard → Datasets → Upload. Give it a clear name (e.g. customer_transactions_q1.csv) — the planner uses this name verbatim when selecting your data.

Create an agent project

Click Data Scientist in the top nav → New project. Fill in a title and a concrete goal:

Using customer_transactions_q1.csv, generate a forensic report that flags columns whose first-digit distribution deviates from Benford's law. Also produce a visualisation of the top 10 numeric column distributions.

The scope gate rejects off-topic / NSFW / gibberish goals at creation time with HTTP 422 and a human-readable reason. No AI quota is consumed.

Review the plan

Within 3-15 seconds a plan appears — numbered steps with tool names, plus a total credit cost estimate at the top. Click Approve plan to begin execution.

ℹSimulations are approved in two parts

A SCADA simulation's cost depends on its scenario pack, so the plan shows it in two parts. Pack preparation: an existing compatible pack is reused at no cost; otherwise the plan shows the one-off cost of preparing a new, reusable pack. Execution: priced from the pack — with an expected cost and a maximum the run cannot exceed. When the pack has to be prepared first, approving the plan covers the preparation only; once the pack is ready the project pauses and shows the execution cost for a separate approval. From the CLI, rady ads decide <id> approve shows that cost and asks you to confirm it; from the SDK, pass max_credits and consent_fingerprint from pending_step_approval["consent"], with the request you reviewed as approval_request_id (the project's waiting_approval_request_id; every decision names what it was made on), to submit_step_approval.

Watch execution

Each step card updates live. Pulsing dot while running, rich result panels on completion (Benford histograms, forensic verdicts, standard tabular benchmark metrics). When a step fails quality checks, self-healing execution proposes a patch plan automatically — you'll see a self-healing badge.

Download the outputs

At complete state: step result panels show inline charts & tables, the Artifacts section lists every file, and the Crypto Trail panel has a Download replay bundle button that exports the full BLAKE3-sealed audit record.

What makes this different

Sealed evidence

Every project emits a cryptographically verifiable sealed evidence bundle. Re-verify months later without trusting our infrastructure.

Typed memory

The planner remembers which tool sequences worked on which domain. Every new project in the same domain benefits from prior runs.

Self-healing

Failed quality gates trigger automatic patch plans, not opaque retries.

Bounded spend

Credit estimate shown before approval. Over-budget projects rejected up-front, never mid-execution.