Understanding meldra

meldra is an agentic data platform for Apache Iceberg. At its center is an AI agent that reads what you're trying to do — in plain English — and carries it out by calling the same tools a data engineer would use by hand: ingest a file, run a query, check a relationship in the graph, sync data across tables. Underneath, everything it touches lands in your own cloud storage as Apache Iceberg — validated, versioned, never silently overwritten. You own the storage. Nothing the Agent does is a black box — every step it takes is one of a fixed, visible set of tools.

Speak meldra

Apache Iceberg

The filing format your tables are saved in. A smarter version of a spreadsheet folder that remembers every past version of itself and never silently loses data.

Catalog

The index that knows which tables exist and where their files live — like a card catalog in a library.

Warehouse

The actual storage location (an S3 bucket, or a local folder in testing) where your table files physically sit.

Namespace

A folder for grouping related tables — e.g. default, finance, hr.

Bronze / Silver / Gold

How "clean" data is. Bronze = raw, as loaded. Silver = cleaned and deduplicated. Gold = ready for business reporting.

Snapshot / Time Travel

Every change is saved as a snapshot. Time Travel lets you view — or restore — the table exactly as it was earlier, useful for "what did this look like before the bad import."

Schema Evolution

Adding, removing, or renaming a column without rebuilding the whole table.

Data Contract

A rule on a table — e.g. "this column can never be empty." Contracts can block bad data from loading at all.

Ingestion

Loading data in, from a CSV file or a connected system like SuccessFactors.

Append / Overwrite / Upsert

Append adds new rows. Overwrite replaces everything. Upsert updates existing rows and adds new ones, matched on a key column.

RBAC

Rules about who can see what. A Business Analyst might see a vendor table with the bank account column blanked out; an Admin sees everything.

Column Masking

Redacting a specific column's real values for certain roles, instead of hiding the whole table.

Query Lab

Where you write SQL, run a Python script, or explore graph connections directly against your tables — the same tools the Agent reaches for, just driven by you instead of a plain-English request.

DuckDB

The engine that runs your SQL queries fast, without a separate database server.

The meldra Agent

An AI agent (built on Claude) with a fixed toolbox: create a table, ingest a CSV, run a SQL query, analyze a graph, sync data to the graph store, query the graph store. Ask it something in plain English — "top 10 vendors by spend in EMEA" — and it decides which tool(s) to call, in what order, runs them, and can use the result of one step to decide the next. That decision loop is what makes it "agentic" rather than a single fixed lookup.

Tool-calling

How the Agent actually does things instead of just talking. Each capability is exposed to it as a discrete, named action it can invoke with specific arguments — the same action you'd trigger by hand in Query Lab, just requested in plain English instead of clicked.

Agent Loop

The cycle the Agent runs per request: read what you asked → decide whether a tool is needed → call it → look at the result → decide the next step or give you a final answer. Capped at a fixed number of steps, so it can't run away indefinitely.

Graph tab

Your data as connections rather than rows — e.g. which vendors share a bank account, useful for spotting fraud rings.

Audit Log

A record of who did what, when, across the platform — including every tool call the Agent makes on your behalf — for compliance and troubleshooting.

Getting Started — First 10 Minutes

1. Register an account (first account on a fresh instance becomes Admin automatically).
2. Verify your email with the code sent to you.
3. Go to Ingest, upload a sample CSV, review the auto-detected schema, and load it.
4. Go to Query Lab, run SELECT * FROM iceberg_table LIMIT 10 against the table you just created.
5. Go to the Agent and ask a plain-English question about the same table — watch it choose a tool and run it.
6. Go to Audit and see that both actions above were logged, including the Agent's tool call.