Hyperparam

HypStack

Own your AI data.

An open observability stack for the AI era.

/ 02The landscape

You already have a data stack.

Datadog Splunk Snowflake Databricks BigQuery Redshift Elastic New Relic CloudWatch Honeycomb Kafka dbt Airflow Looker Tableau

…and none of it was designed for AI.

/ 03The shift

The shape of data is changing.

A new class of applications

Codexclient-side

Claude Codeterminal

Cursordesktop IDE

Claude Desktopdesktop app

Copilotin-editor

your app + LLMbrowser / mobile

Compute moves toward the client.

…producing piles of text

14:23:01 user      refactor the auth middleware to use jose
14:23:02 system    retrieved 12 docs · 38 KB context
14:23:03 tool      grep("jsonwebtoken", src/) → 7 hits
14:23:04 tool      read_file(auth/middleware.ts)
14:23:05 tool      read_file(auth/verify.ts)
14:23:06 tool      edit(auth/middleware.ts) ✓
14:23:07 tool      edit(auth/verify.ts) ✓
14:23:08 tool      edit(package.json) ✓
14:23:09 tool      run_tests() → 3 passed
14:23:10 tool      typecheck() → clean
14:23:12 assistant Here's the refactor — I switched to jose…
14:23:14 user      looks good. add a test for expired tokens
14:23:16 tool      read_file(auth/middleware.test.ts)
14:23:18 tool      edit(auth/middleware.test.ts) ✓
14:23:21 tool      run_tests() → 4 passed
14:23:23 assistant Added an expired-token case. Ready for review.
14:23:23 usage     in=28,140 · out=6,820 · ~118 KB session

Megabytes of text per session, not bytes per request.

A new shape of data. A new shape of stack.

The real shift isn't "more logs" — it's a different shape of data entirely. Two things are changing at once.

First: a whole new class of applications. Codex, Claude Code, Cursor, Claude Desktop, Copilot, plus every product team bolting an LLM into their own app. These run on the client — terminal, IDE, desktop, browser — not behind your load balancer.

Second: what they emit. Not "GET /users 200 23ms." A single agent turn is a 50 KB document — prompt, retrieved context, tool calls, tool responses, completion, token counts. A real session is hundreds of those. That's megabytes of structured text per user per hour.

Datadog wants 200-byte structured events. Snowflake wants normalized rows. Neither was designed for this shape.

/ 04The squeeze

Two stacks. Neither one fits.

Observability vendors

billed per GB ingested
30-day retention by default
structured for metrics, not text
your data, their walled garden

Warehouses

great at SQL, bad at nested LLM payloads
compute is rented, not yours
governance lives in the vendor
painful to give to a notebook or agent

Two stacks, neither built for AI workloads —
both billing you for the privilege.

/ 05Today's answer

Yet another vendor.

LLM-obsprompt traces

Eval SaaSscoring runs

Agent-tracetool calls

RAG analyticsretrieval logs

Prompt mgmtversion + diff

A new pane of glass for every new shape of log —
each one a copy of your data, behind their API.

/ 06The reframe

What if you owned your data?

Open format. Your bucket. Any tool that speaks it.

SourceAI Agents

OTel

StoreIceberg

AnalyzeHyperparam

Collect with the open standard. Store in the open format.
Analyze with the open client. No proprietary middle.

/ 07The loop

Close the loop on your agents.

Company data shapes your AI → AI emits logs → Collectivus centralizes → Iceberg on your S3 stores → Hyperparam finds what works → updates flow back into company data.

The whole thing is a feedback loop. Sources on the left — agents, copilots, workforce machines, production services — emit logs. Collectivus, our OTel-based collector, ships to every machine via your existing MDM and aggregates everything into a single Iceberg table on your own S3 bucket.

Hyperparam reads that Iceberg table directly using hyparquet, icebird, squirreling, and parquetindex (full-text search). What it surfaces: where your agents are working well, where they're struggling, which prompts get retried, which tools fail.

The output isn't a dashboard — it's an update to your knowledge base. The CLAUDE.md, prompts, tool definitions, standards, and style guides your team already maintains as markdown. Hyperparam tells you what to change. The change goes back to the agents. The agents get better. The cycle repeats.

That's the pitch: not "buy another logging tool." It's "close the loop between your agents and your knowledge."

/ 08 · Layer 01Collection

Collect from every surface.

Capture every developer, system, user, and agent — without writing a custom SDK for each one.

Collectivus — our OTel collector, deployable across your fleet via MDM
OpenTelemetry under the hood — open standard for traces, metrics, logs
Workforce coverage — laptops, dev tools, agents, production services
Agent & LLM instrumentation — prompts, tools, retrieval, evals
Schema-on-write — structured payloads, not blob dumps

We help you instrument. The data is already yours.

/ 09 · Layer 02Storage

Store in your bucket.

Your account. Open format. Read by anything that speaks Iceberg.

Apache Iceberg on S3 — the format every warehouse is converging on
Cheap durable storage — pennies per GB, infinite retention
Schema evolution & time travel — built into the format
Compaction & partitioning — managed for you, not by you
Catalog of your choice — Polaris, Lakekeeper, Glue, Nessie

Snowflake reads it. Databricks reads it. So does Hyperparam.

/ 10 · Layer 03Analysis

Analyze from anywhere.

A browser-native debugger for chat logs, agent traces, and evals — no cluster, no SQL endpoint, no ingestion.

Pure-JS Iceberg + Parquet client — millions of traces in a tab
Joins across sources — Iceberg, GitHub, S3, existing warehouses
SQL, search, geospatial, ML — hyparquet, icebird, squirreling, parquetindex
Read-only, local-first — credentials stay in the browser
Embeddable — VS Code, Claude Desktop, agents themselves

One client. Every source you already have.

/ 11The workflow

Explore. Surface. Improve.

Not another dashboard for a metric you already track. A debugger for the new shape of AI data.

01Explore

Browse millions of rows in a tab — JSONL, Parquet, Iceberg, no ingest
AI-assisted search across prompts, tools, retrievals, completions
Join with code — repos, PRs, issues, your warehouse

02Surface

AI-generated labels classify failure modes at scale
Find the 1% that's broken — filter, cluster, slice
Token-burn hotspots, retry loops, tool failures

03Improve

Validate prompt + tool changes against real traces
Update the knowledge base — CLAUDE.md, prompts, skills
Ship — close the loop back to your agents

Ground every change in real trace data — not speculation.

This is the actual workflow inside Hyperparam — what you do once the data is in Iceberg.

Explore: open a 2.3M-trace dataset in a browser tab in under a minute. AI-assisted search lets you find the conversation, retry, or tool call you're looking for without writing SQL. Join with GitHub means correlating an agent's actions with the code it was editing.

Surface: this is the part teams underinvest in. Hyperparam generates label columns at scale ("did this conversation succeed?" "was the tool call necessary?") so you can filter, cluster, and slice millions of rows down to the 1% that's actually broken — token-burn hotspots, retry loops, tool failures.

Improve: a prompt change isn't a vibe — it's a hypothesis you validate against the traces you already have. The change goes back into the knowledge base — CLAUDE.md, prompts, skills — and from there, back to the agents. Loop closed.

/ 12How a query flows

Six steps. Zero servers in between.

	step	where it runs
01	Agent / copilot / dashboard issues a query	`client`
02	Iceberg catalog: resolve table	`your catalog`
03	Read manifest list, prune partitions	`client`
04	Read manifests, prune data files by stats	`client`
05	Fetch only the Parquet byte ranges needed	`your S3`
06	Decode, filter, render	`client`

Zero servers between the user and their data.

/ 13Outcomes

What this enables.

Full-fidelity AI observability — prompts, tools, retrieval, evals, all queryable
Infinite retention at object-storage prices, not Datadog prices
Agents that read their own traces direct from S3, no SQL endpoint
Notebook-grade analysis for ML engineers without warehouse credits
One copy of the data — shared across ops, analytics, ML, product
Compliance that doesn't fight you — your IAM, your VPC, your region

Same logs. Same format. No vendor in the middle.

/ 14Engagement

Where we come in.

You bring

your apps, agents, and services
your S3 / GCS / Azure account
your existing warehouse (optional)
your security & compliance posture

We bring

OTel + MDM instrumentation patterns
opinionated schemas for AI workloads
collector & Iceberg deployment
Hyperparam analysis client + support

Weeks to first value. No data leaves your account.

/ 15Takeaway

The AI era doesn't need a new vendor.
It needs a stack you can own.

Open collection. Open storage. Open analysis. Every layer MIT or Apache 2.0. Every layer replaceable — including us.

Next time someone offers a dashboard for your AI logs,
ask whether you can just own the data instead.

Hyperparam

HypStack

Own your AI data.

Open observability for the AI era.

Cost shape

Datadog logs: ~$0.10–$2.50 per GB ingested, 15–30 day retention default
Splunk: ~$0.50–$5 per GB ingested, premium for retention
S3 standard: ~$0.023 per GB-month, infinite retention
S3 Glacier IA: ~$0.004 per GB-month for cold AI traces

Two orders of magnitude on storage. Compute moves to the client.

Who owns what

asset	traditional	HypStack
raw logs	vendor cloud	your S3
schema	vendor-defined	open Iceberg
query engine	vendor	JS client / your warehouse
retention policy	vendor pricing tier	S3 lifecycle
access control	vendor IAM	your IAM

Own your AI data.

You already have a data stack.

The shape of data is changing.

Two stacks. Neither one fits.

Observability vendors

Warehouses

Yet another vendor.

What if you owned your data?

Close the loop on your agents.

Collect from every surface.

Store in your bucket.

Analyze from anywhere.

Explore. Surface. Improve.

01Explore

02Surface

03Improve

Six steps. Zero servers in between.

What this enables.

Where we come in.

You bring

We bring

The AI era doesn't need a new vendor.It needs a stack you can own.

Own your AI data.

Cost shape

Who owns what

The AI era doesn't need a new vendor.
It needs a stack you can own.