Introduction

Pramen (Czech for spring — a source of water) is a data movement and transformation runtime built around three commitments:

1. Columnar and bounded. Data moves as Apache Arrow record batches through bounded channels. A slow destination applies backpressure all the way to the source, so peak memory does not grow with input size.
2. Governed LLM enrichment. Semantic transforms (ai.extract, ai.classify, ai.generate) are schema-bound operations with hard budgets, strict output validation, and a durable, content-addressed inference ledger. Completed model calls are never re-billed — not after a crash, not on a re-run, not on a partial replay.
3. Radically simple operations. One static binary with zero native driver dependencies. A pipeline is one YAML file. The v1 promise is measurable: download the binary, write the file, and have enriched rows in PostgreSQL in under ten minutes.

Who it is for

Platform and data teams moving data from object storage into operational databases — especially when some columns need a language model to exist at all (classification, extraction from free text, normalization), and when that inference has to be auditable, budgeted, and restart-safe.

What it is not

• Not a warehouse. Pramen feeds databases; it does not replace them. If your data already lives in a warehouse with AI SQL functions and the results stay there, use those.
• Not a streaming platform. There are no topics, consumer groups, or cluster coordinators. Pramen scales through independent workers first.
• Not an agent framework. Semantic transforms have fixed inputs, fixed instructions, typed outputs, and no tools or loops. That constraint is what makes governance enforceable.

Current status

Workspace 0.2.0 ships the Phase 1 verticals plus Phase 2 Group X1: deterministic and governed-AI pipelines, Azure/GCS sources, WASM (local + OCI-by-digest), ai.generate, and optional Postgres ledger/checkpoint backends. Live cloud acceptance and Phase 2 Group X2 remain open. The status and roadmap page tracks exactly what works now, and every performance claim on this site links to a measured result.