Release 0.2.4 · runs in your perimeter

The lakehouse that
moves at the speed of
intent.

blynk is an AI-operated lakehouse control plane. You describe the outcome. The agent drafts a validated plan. You approve. Temporal executes it against governed Iceberg — and nothing touches production without a human saying yes.

See the loop
plan-first writes
0%
pinned engines
0
data leaving your VPC
0

Try it

Chat intent → validated plan → human approve → Temporal → governed Iceberg

This is the actual product loop, running as a mock-up. Pick an intent and watch the control plane refuse to guess.

Project · retail-analytics
autonomy: plan-first catalog lake.sales

Project bound to lake.sales · federated: pg_prod, mysql_billing. Every guarded write waits for your approval.

Suggested intents

Guarded writes — ingest, write SQL, provisioning, drops — always stop for Approve. Reversible write_safe actions can run autonomously, journaled with a one-click undo.

Why we build blynk

Data platforms got powerful. They also got heavy.

Teams still hand-stitch catalogs, query engines, pipelines, identity and AI tooling — then hope governance holds under real production load. Handing an LLM the keys is faster and far worse.

01

Humans approve change

The agent proposes; operators decide. Approval is a product primitive, not a policy document.

02

Governance ships with the plane

Identity, catalog grants and per-query policy are enforced by the engines, not bolted on later.

03

Private by default

Local or private cloud. The model, the metadata and the bytes stay inside your perimeter.

04

Open foundations

Iceberg, Polaris, Trino, SeaTunnel, Temporal — pinned, upgradable, never a black box.

Platform

One console for the whole lakehouse

Projects that remember

Durable, named chat workspaces bound to a catalog and namespace. Messages, plans, runs and decisions live together, so the "why" survives long after the query.

  • Plan versioning with edit-in-place
  • Temporal deep links from every run
  • Grounded in real information_schema

Governed Iceberg

SeaTunnel writes into Apache Iceberg tables cataloged by Polaris, with per-user principals.

Federated sources

Attach Postgres, MySQL and JDBC catalogs to Trino, then plan across them in one prompt.

Local AI, no exfiltration

llama.cpp drafts plans, writes SQL and diagnoses failures — on hardware you control.

Approval as infrastructure

Every guarded mutation resolves to a plan a person signed. Reversible actions are journaled with the exact operation that undoes them.

Serve & explore

Trino Gateway routing plus Superset dashboards, wired to the same governed tables.

Architecture

Hexagonal core, pinned engines, one perimeter

The control plane talks to ports. Engines are adapters — swappable, version-pinned, and never hardcoded into business logic.

SOURCES CONTROL PLANE EXECUTION GOVERNED LAKE Postgres · MySQL S3 · CSV · Parquet Kafka · APIs Project chat Plan (validated) Human approve llama.cpp · sqlguard · ports Temporal SeaTunnel Zeta Trino Iceberg tables Polaris catalog Superset · Gateway Keycloak OIDC · OPA per-query authorization · access mappings · audit journal

Pinned in deploy/versions.yaml

no :latest on the release path
  • Trino479
  • SeaTunnel2.3.11
  • Polaris1.2.0
  • Temporal1.27.2
  • Keycloak26.1
  • OPA1.19.0

Governance

The guardrails are the product

An agent that can reach production is only safe if the blast radius is defined by the platform, not by prompt discipline.

  • Write SQL is validated, not trusted. sqlguard.ValidateWrite allowlists verbs and rejects federated sinks before a plan can be approved.
  • Authorization is enforced per query. Keycloak identity resolves to engine principals and catalog grants; Trino asks OPA on every statement.
  • Tool authority is an intersection. What the agent may do = the caller's roles ∩ the project's autonomy setting. Never more.
  • Reversible writes are journaled. Each write_safe action records the operation that undoes it, so autonomy stays recoverable.

Put AI in the lakehouse loop —
and keep the keys.

Deploy the control plane into your own Kubernetes or private cloud. Bring your catalog, your identity provider, and your rules.

Requests go to this deployment’s API (/api/v1/public/access-requests).