Nicholas Lippis

Software Engineer, Observability, Streaming Data Processing, Automated Reasoning, Distributed Systems

[Grepr @grepr_ai]

Three AI labs, three weeks, the same basic failure: OpenAI (July 21), Anthropic (July 30), and Meta (early August) each disclosed models reaching the open internet from inside sealed-off cybersecurity evals and touching real companies' production systems, mostly using ordinary techniques like weak passwords and SQL injection. Two of the three cases traced to the exact same third-party evaluator misconfiguration.

We're building Proactive Anomaly Detection because catching this shouldn't depend on a lab happening to review the right transcript.

The idea: no human team, and no static rules engine, can watch every log for every pattern worth catching. So we're designing our system to continuously extract the recurring shapes of an organization's full telemetry stream, in-line, no sampling, and have an LLM agent reason over those shapes: what's familiar, what's new, what just started behaving differently. Where it's confident, it turns that understanding into a standing detection automatically, no human has to write the rule first.

That's correlation and analysis at a scale no human security team can do by hand, and no rules-based tool was built to do.

We're building and validating this with a small group of design partners now. Full breakdown of the three incidents and how the architecture maps to each: [link to blog]

Early access is open for design partners.

https://t.co/dJH6U3eIAV

[Grepr @grepr_ai]

A global retailer with 85K locations just saved >$1.5M on their New Relic bill.

They tested Grepr on full production traffic and cut bytes to New Relic 78%, taking 75% off total cost. Dashboards and alerts never changed.

Read the story: https://t.co/mZXgYTP5Y6

[Jad Naous @jadtnaous]

What if you could have an AI agent monitoring every log line, and you didn’t have to worry about context windows and cost?

But you can’t run inference on every log line and maintain that much context.

That’s the problem we’re solving at @grepr_ai.

I wrote about it here:
https://t.co/Shwj4sYg9z

[Grepr @grepr_ai]

Your logs saw the outage coming. Your alerts didn't.
The failures that take systems down are the ones no one thought to alert on. The signal was in your logs the whole time.

How Grepr brings AI to every log without inference on every log: https://t.co/H95sCPUJhk

[Grepr @grepr_ai]

@grepr_ai support for distributed traces is now GA.
Path-based tail sampling keeps every unique execution path and drops the redundant traffic, so you cut trace volume without losing the traces you need.
FOSSA cut trace volume 90%.

https://t.co/II8hnD1ORw

[Grepr @grepr_ai]

With Grepr, @envoy cut log volume by more than 90% and flattened a cost curve that had been compounding for years, without changing a single dashboard, alert, or engineering habit.

Read How: https://t.co/xEMeNDN3tI

[Grepr @grepr_ai]

Grepr's free tier is live and perpetual. No credit card, unlimited users, 4 TB logs and 8 TB traces a month, two pipelines. Same pattern engine as Pro. One config line, patterns in 30 minutes, all on your own data.

[Jad Naous @jadtnaous]

@Intellyx recognized @grepr_ai with a 2026 Digital Innovator Award, and what I am proud of is the reason: we shred the bill while keeping full fidelity. https://t.co/9V6DAzCC99

[Grepr @grepr_ai]

Most observability stacks collect everything, then ask engineers to find signal in the wreckage.

@jadtnaous sat down with @mirko_novakovic at @dash0hq to talk about why that model broke a long time ago, and what comes next.

https://t.co/vUPl8cLLKe

[Grepr @grepr_ai]

Grepr is sponsoring Observability Summit NA, May 21-22 in Minneapolis.

Find the team at Table 16.

https://t.co/1D8c5zrT8f

[Jad Naous @jadtnaous]

New piece on TFiR: observability at scale has become a configuration management problem.

95% of log messages are repeated patterns. Most teams are paying to ingest the noise around the signal.

Detection times drop from hours to minutes when the pipeline does the configuration toil for you.

https://t.co/wSvXV1b9le

[Grepr @grepr_ai]

132GB of data served to users. 38 petabytes of telemetry to do it.

That's the actual ratio. And it's why the Datadog bill keeps going up while nobody's resolving incidents any faster.

We got into the math: https://t.co/wrEvoNBIn7

[Grepr @grepr_ai]

HIPAA wants six years of audit logs. Your observability platform wants six figures a month to store them.
The fix is a two-tier architecture: reduced signal to your platform, raw events to S3 with Parquet and Iceberg. Query years of history in seconds. https://t.co/oiNsyHhaYN

[Grepr @grepr_ai]

The Grepr team will be at #O11ySummit next month in Minneapolis.

Still need to grab your ticket? Head here: https://t.co/Ds0P9xQwtL

We'll see you at table T16!

[Grepr @grepr_ai]

Most observability platforms charge you to keep your own logs.
The architecture for getting out from under that, Parquet + Iceberg + a query layer that actually works:
https://t.co/GozEosBS3E

[Grepr @grepr_ai]

Here's what we learned at #Kubecon last month:

#AI workloads are generating #telemetry at volumes that most budgets were never built to handle, and teams feel it every time they open an invoice.

#Grepr COO @Johnymkim wrote a recap here: https://t.co/060Zk4PB4h

[Grepr @grepr_ai]

Your healthcheck logs are probably 15-40% of your total log volume. Millions of identical lines per day, billed at the same rate as the logs that actually matter. Here's how to fix that: https://t.co/1edxlE4rI0

[Grepr @grepr_ai]

keeping every log "just in case" isn't a monitoring strategy... it's what happens when retrieval is broken.

https://t.co/DIEPZDB8Gj

[Grepr @grepr_ai]

Application logs and APM traces aren't the same data type. Here's how they differ, why it matters for your observability bill, and when to use each.
https://t.co/HIzorCjTlF

[Grepr @grepr_ai]

Our CEO @jadtnaous talked about the rising costs of infrastructure and it's autonomous future with @NickLippis. Between a mountain of noise and a rapidly changing environment, SREs can't catch up...