What is Radhflow?

You have a CSV of leads. You want to score them, filter the top 20%, and push them to your CRM. Every week, automatically.

You could write a Python script. You could wire up Zapier. You could ask an LLM to do it every time it runs — burning tokens and getting slightly different results each time.

Or you could describe what you want, let AI write the pipeline once, and run deterministic code forever.

That is Radhflow.

The compiled agent model

Radhflow is a compiled agent platform. You describe a data pipeline in plain language or YAML. AI generates typed, deterministic code — SQL transforms, API connectors, shell scripts. That code runs without any LLM in the loop. Zero token cost at runtime. Same input, same output, every time.

 You ──describe──> +------------+    +----------------+
                   | Conductor  |--->| Code Agents    |
                   | (AI)       |    | (AI, one-time) |
                   +------------+    +-------+--------+
                                             | generates
                                             v
                   +-------------------------------------+
                   | Pipeline (flow.yaml + scripts)      |
                   |                                     |
                   |  +------+   +------+   +------+    |
                   |  | Node |-->| Node |-->| Node |    |
                   |  +------+   +------+   +------+    |
                   +-------------------------------------+
                                             | executes
                                             v
                   +-------------------------------------+
                   | Deterministic Runtime               |
                   | DuckDB . SQL . HTTP . Shell          |
                   | No LLM. No tokens. No surprises.    |
                   +-------------------------------------+

AI touches your pipeline once — at creation time. After that, it is plain code.

The key insight

AI is expensive and non-deterministic. Use it once to write the pipeline, then run deterministic code forever. You get AI’s ability to understand what you want without AI’s cost and unpredictability at runtime.

How it works

A pipeline is a flow.yaml file. It describes a directed graph of nodes connected by edges. Data flows from sources through transforms to outputs, and every connection is typed.

flow.yaml

nodes:
  read-leads:
    type: source
    op: file.read_csv
    params:
      path: leads.csv
    outputs:
      leads:
        type: Table

  filter-top:
    type: deterministic
    op: sql.query
    params:
      query: "SELECT * FROM leads WHERE score >= 80"
    inputs:
      leads:
        type: Table
        from: ref(read-leads.leads)
    outputs:
      qualified:
        type: Table

  write-output:
    type: deterministic
    op: file.write_json
    params:
      path: qualified-leads.json
    inputs:
      qualified:
        type: Table
        from: ref(filter-top.qualified)

Three nodes. CSV in, SQL filter, JSON out. Every connection carries a type contract. Schema mismatches are caught before any code runs.

Core properties

• Deterministic execution. No LLM at runtime. Pipelines produce identical output for identical input.
• Four data types. Value (scalar), Record (single object), Table (rows), Stream (unbounded rows). Nothing else needed.
• NDJSON interchange. Tables are .ndjson files with .schema.json companions. Human-readable, diffable, universal.
• DuckDB transforms. SQL is the default language for filtering, joining, aggregating. No custom DSL.
• Git-native. Branch a pipeline. Diff it. Review it in a PR. Merge it.
• Local-first. Runs on your laptop. Cloud deployment optional. EU infrastructure when you need it.

Who it’s for

Role	Use case
Data engineer	Typed, reproducible ETL pipelines versioned in Git.
Marketing coordinator	Lead scoring, campaign reporting, CRM sync — no engineering tickets.
Ops team	Automate invoicing, inventory, reporting — own the pipeline forever.
AI agents	Structured execution backend via MCP server.
Freelancer	Client data workflows that run locally without cloud subscriptions.

What Radhflow is NOT

Not a chatbot. There is no conversational AI at runtime. AI generates code once. That code is static and deterministic.

Not an agent framework. There is no autonomous agent loop running on every execution. The AI’s job ends when the pipeline is compiled.

Not a no-code tool. Radhflow generates real code. You can read, edit, and extend every file. The visual canvas is a view into YAML, not a replacement for it.

Caution

Radhflow is not a data warehouse. It moves and transforms data. Use it alongside Postgres, BigQuery, or whatever stores your data.