Hyperparam provides a platform that records and indexes all AI‑related traffic—prompts, tool calls, and session data—so teams can query their usage directly from storage. It lets users break down token spend by developer, repository, and workflow, and discover usage patterns across agents with keyword, semantic, and SQL searches that run in the browser against the raw log files. This visibility turns scattered AI logs into actionable insights for cost control and productivity.
Funding
Funding not disclosed
Founders
Product
Problem
AI teams generate massive amounts of token‑rich logs from agents, chatbots, and coding tools, but these logs are scattered across storage systems and stored as large, nested JSONL or Parquet files that traditional analytics tools cannot read effectively. Without a way to query the actual content, teams cannot answer basic questions about token spend, prompt effectiveness, or failure patterns.
Solution
Hyperparam provides a browser‑native platform that reads agent and chat logs directly from object storage (S3, GCS, Azure, Hugging Face, Iceberg, etc.) without requiring data ingestion or copying. Users can search logs by keyword, semantic meaning, or SQL, and invoke an integrated AI assistant to generate derived columns, classify failures, and suggest prompt fixes. The system streams only the bytes needed for each query, enabling interactive analysis of multi‑gigabyte datasets. By combining natural‑language queries, SQL views, and cross‑source joins, Hyperparam lets teams surface cost drivers, debug tool‑call failures, and improve prompts directly on production traces.
Target Audience
Primary users are AI product and platform teams, agent/LLM developers, and ML engineers who need to debug production agent traces, chatbot histories, and tool‑call logs at scale.
Features
- Direct, lazy streaming of multi‑gigabyte Parquet and JSONL logs from cloud buckets via HTTP range requests, keeping billions of rows responsive in the browser
- Integrated AI agent that can be prompted in plain language to add derived columns, classify errors, score quality, and generate remediation suggestions
- Full‑text, semantic, and SQL search across nested conversation structures without flattening the data first
- Cross‑source joins that combine logs with code repositories, issues, or other datasets, enabling correlation of agent behavior with underlying code changes
- Support for a wide range of data sources including local files, S3, GCS, Azure Blob, Hugging Face, and Apache Iceberg tables
- Export of transformed datasets in Parquet, JSONL, or CSV for downstream fine‑tuning, evaluation, or reporting
- Reusable “skills” that capture analysis workflows to be rerun automatically on new log batches