Antidha
Checking the demo pipeline…

Know why your real-time and batch numbers disagree.

Antidha is a copilot for teams that run a streaming path and a batch path over the same data. It reads live pipeline state, does the arithmetic in code, and has Claude explain what diverged, why, and what to check next.

Ask the live demo See the reference pipeline

Early-stage prototype. The demo runs on a reference pipeline, not on customer data.

How data flows through the pipeline Source data feeds a fast, approximate speed layer and a scheduled, tested batch layer. A serving layer merges both, and the Claude copilot reads all three to explain differences. Source Speed layer every 2 s · approximate Batch layer every 5 min · dbt-tested λ Serving Claude explains the gap

The problem

Two paths, two answers, one on-call engineer.

A Lambda-style setup computes the same figures twice: fast and approximate on the streaming path, slow and authoritative on the batch path. Sooner or later the dashboard and the report disagree.

Is it a failed batch run, a stale layer, a dbt test that stopped passing, or just rounding? Answering means reading run history, test results, logs and two data stores side by side. That is the work this tool does for you.

Live demo

Ask the demo pipeline.

A reference pipeline is running right now at portfolio.herrise.cloud. The briefing below is computed from its live data; ask Claude anything about how it is behaving.

Pipeline briefing

Loading

Reading the pipeline…

    Ask the pipeline

    0 / 400

    The speed-layer prices on the reference pipeline are simulated, so this is demo data, not market data, and the copilot gives no trading advice. Your question and a snapshot of the pipeline's state are sent to the Anthropic API to produce an answer. Answers can be wrong: check the evidence.

    How it works

    The model explains. Code does the arithmetic.

    1. Read

      Pulls status, run history, dbt test results, logs and both the speed and batch views from the pipeline's API, in parallel, behind a short cache.

    2. Compute

      Plain code works out the gaps between layers, how old each value is, which gaps matter, and whether a gap is only rounding. These numbers are never left to the model.

    3. Explain

      Claude answers only from that evidence, separates what it observed from what it infers, says what it cannot tell, and shows the evidence next to every answer.

    • Per-visitor rate limits
    • Daily spend cap
    • Log text is treated as data, never as instructions
    • No trading advice

    Where things stand

    An honest status.

    Working today

    • A live reference pipeline: Redis speed layer, DuckDB and dbt batch layer with automated data tests, FastAPI serving layer.
    • A copilot that writes a status briefing and answers questions from that pipeline's live data.
    • The reference pipeline's source is public on GitHub.

    Not yet

    • Connecting your own pipeline. The demo reads one specific API.
    • Accounts, authentication and per-team data isolation.
    • Customers. This is an early prototype.

    Run a streaming and batch setup and want to try this on it? hello@antidha.com