ClickHouse AgentArena

Crown the winner.
Prove the cost.

In 90 minutes you run your own contest — six models across three prompt strategies, answering real business questions in SQL over ClickHouse — and pick what to ship on evidence: cost per correct answer, not a public leaderboard.

90 minutes · start to finish18 configs gradedLangfuse from step 0$0 on trial credits

Not a public leaderboard. Your own contest.

Picking a model is the first consequential decision in an agent application, and the one most often guessed. Guess high and you pay frontier prices for capability the task never needed. Guess low and you ship something that quietly gets your real workload wrong.

So you run the contest yourself. A grid of six models across three prompt strategies answers a ground-truthed set of business questions against a live ClickHouse dataset.

Every answer is graded on execution accuracy — did the query return the right result, not the right-looking SQL — and every configuration is ranked by cost per correct answer. That is quality per dollar for your task, which is the only number that can actually settle the argument.

Langfuse is wired in from module 00, not bolted on at the end: it runs the contest as experiments, grades it with a code evaluator and an LLM judge, stores every trace, and follows the winner into production.

What you walk away with

Everything you'll have running by the end.

On your own ClickHouse Cloud trial, your own Langfuse project, and one OpenRouter key.

01

A contest, run and settled

Eighteen model × prompt configurations graded against your golden questions and ranked by cost per correct answer.

02

Langfuse experiments and evaluators

A correctness evaluator and an LLM judge running server-side in your own Langfuse project, scoring every run.

03

Grading you can defend

Answers compared to cached golden result sets — column position, float rounding, row-set normalisation — so a lucky-looking query cannot pass.

04

A continuous-improvement loop

Reviewed answers become new golden questions and the grid re-runs — the whole loop visible end to end in Langfuse.

05

The winner in production

The winning config live behind POST /ask, traced per session, with 👍/👎 written back as Langfuse scores.

+

The whole repo

Harness, agent, read-only SQL guard, serving API and arena UI — yours to point at your own data.

The route · ~90 min hands-on

Five stops, in order.

One flow, start to finish: pick a model on evidence, understand it, improve it, then watch it in production.

  1. 00

    Setup

    20 min

    Connect OpenRouter, ClickHouse Cloud, and Langfuse — tracing wired in before any model is chosen — and seed the arena dataset.

  2. 01

    Select the base model

    20 min

    Run the model × prompt grid as Langfuse experiments and crown a winner by cost per correct answer.

  3. 02

    Measure quality

    15 min

    Go past “it won”: per-tier accuracy, the llm_judge signal, and individual traces from prompt to generated SQL.

  4. 03

    Improve continuously

    20 min

    Turn reviewed answers into new golden questions through the annotation queue, then re-run the grid against them.

  5. 04

    Release to production

    15 min

    Serve the winning config behind a real API and watch production traces, sessions, cost, and 👍/👎 feedback.

Self-paced by default. Every module names its starting checkpoint and ends with a verification you can check yourself, so nobody is stranded by the pace of the room.

Before you join

Who it's for, and what to bring.

Who it's for audience

  • Engineers and architects about to pick a model for an agent feature
  • Anyone who has been asked “why this model?” and wants a defensible answer
  • Teams standing up LLM evaluation for the first time
  • Comfortable with SQL and a terminal — no ML background needed

What to bring prerequisites

  • Python 3.11 and Node 18+ on your laptop
  • A ClickHouse Cloud service with trial credits
  • A Langfuse Cloud project
  • An OpenRouter API key — one OpenAI-compatible endpoint fronts every model in the roster, so there are no per-provider credentials to manage

How it runs format

  • 100% hands-on across five modules — every command and query is copy-paste
  • Fully self-paced, with an instructor track mirroring each module for facilitators
  • Grading runs in your Langfuse project, so the evidence stays yours
  • The roster is deliberately low-cost — NL→SQL does not need frontier models

Ready to run the contest?

Bring a ClickHouse Cloud trial, a Langfuse project, and an OpenRouter key. Leave knowing which model to ship, what it costs per correct answer, and how to prove it again next quarter.

18Model × prompt configurations graded in one run.
$0Trial credits and a low-cost roster cover the session.
YoursThe harness is yours to point at your own questions.