About Hyperparam

The browser-native debugger for agent and chat logs

Agents, coding tools, and chatbots generate quadrillions of tokens of conversational and tool-call data every year. Buried in those logs is everything teams need to understand what their AI is actually doing in production: where conversations go off the rails, where tool calls fail, which prompts burn tokens, and how each release shifts behavior.

But existing tools are not built for this. Observability dashboards aggregate, sample, and flatten. Jupyter notebooks and warehouse SQL choke on multi-gigabyte JSONL of nested conversations. No one can read through millions of rows of text, and the most important signal often lives in a long-tail 1% you would never spot with sampling.

A debugger for agents and chatbots, built for the browser

Hyperparam reads agent traces, coding-tool transcripts, and chatbot histories straight from where they already live: local files, S3, GCS, Azure Blob, Hugging Face, Iceberg tables, even GitHub repos and issues. It pairs them with an AI agent that analyzes the logs alongside you. Drill into nested traces, generate derived columns at scale, build SQL views that join across sources (so agent behavior can be correlated with the code, issues, or prompts that drove it), and save reusable analyses as skills. Because everything runs in the browser, teams can iterate on real production traces without moving the data.

As AI systems keep scaling, the work of improving them depends on what their logs reveal. Hyperparam exists to make that revealing fast, repeatable, and grounded in the actual data: explore, surface issues, improve the prompts, tools, and routing, then ship the fix and watch the next batch of logs.

An open stack for AI observability

The bigger picture: every company already has a data stack, and none of it was designed for AI. Datadog and Splunk are billed per GB ingested with short retention. Snowflake and Databricks are great at SQL but bad at nested LLM payloads, and the compute is rented. The usual answer is to bolt on yet another vendor (LLM observability, evals, agent tracing, prompt management), each one a copy of your data behind their API.

We think the answer is a stack you own end to end. We call it HypStack: open collection with Collectivus on top of OpenTelemetry, open storage in Apache Iceberg on your own object storage, and open analysis with Hyperparam directly in the browser. The stack is MIT licensed and dependency-free, so there is one license to read and no supply chain to audit. Your prompts and traces stay in your bucket, in your IAM, in your region, with no vendor copy and no per-GB ingest fees.

We are hiring! >>