Mendral Fed Terabytes of CI Logs to an LLM and Got 99.9% Data Reduction

Last week I was debugging a flaky CI pipeline at 2am, scrolling through thousands of build logs like a digital archaeologist. Then I stumbled across Mendral's blog post about feeding terabytes of CI logs directly into an LLM to generate SQL queries. My first thought? This sounds insane. My second? This might be genius.

Mendral, an AI startup focused on developer workflows, basically took the nuclear option for log analysis. Instead of the usual dance of preprocessing, schema design, and manual query writing, they dumped massive unstructured CI logs into a large language model and asked it to write SQL in plain English.

The Hacker News crowd was intrigued—164 points and 89 comments worth of intrigue. But buried in those discussions were some fascinating concerns about "costs associated with each investigation" and the complexity of "tuning orchestration between agents."

The Numbers Don't Lie

This isn't just a cute demo. Recent research shows LLMs can achieve 99.9% data volume reduction on terabyte-scale logs while generating causal graphs and automated reports. One study processed logs by:

• Consolidating files chronologically
• Creating templates for common patterns
• Using CPU-based LLM inference for classification
• Building time-series analysis entirely without GPUs

But here's where it gets interesting: while GPT-4o can handle 1.3M-row datasets effectively with proper schema awareness and few-shot prompting, the debugging story is much darker.

That's... not great. Especially when you're dealing with production CI logs where a missed pattern could mean missing the root cause of your 3am pager alert.

The Real-World Reality Check

I've seen this pattern before in AI tooling. The happy path demos are magical—natural language queries turning into perfect SQL, automated insights, beautiful dashboards. Then you hit production complexity.

The challenges are predictable but brutal:

• Cost explosion: Multi-agent orchestration isn't cheap
• Schema confusion: LLMs struggle with complex joins and business logic
• Context loss: Multi-turn corrections often get "lost" in conversation
• Over-confidence: Wrong answers delivered with supreme confidence

One benchmark found LLMs consistently misinterpret business concepts like "active customer"—imagine that error propagating through your CI analytics.

But Maybe That's Missing the Point

Here's what's actually interesting about Mendral's approach: they're not trying to replace human judgment. They're attacking the cognitive load problem that every SRE knows intimately.

When you're staring at terabytes of logs, the bottleneck isn't writing perfect SQL—it's figuring out what questions to ask. An LLM that can quickly surface patterns, generate hypotheses, and create starting-point queries could be transformative, even with a 60% accuracy rate.

Think about it: if an AI agent can spot that mysterious July 4th performance degradation pattern in your logs and generate five different SQL queries to investigate it, who cares if two of them have syntax errors?

The Splunk Disruption Question

This could genuinely threaten traditional log analysis tools. Splunk, ELK Stack, and similar platforms built their moats on query languages and preprocessing pipelines. If LLMs can bridge that gap with natural language, the competitive landscape shifts dramatically.

But only if they can solve the reliability problem. Enterprise customers won't tolerate 32% accuracy on semantic queries, no matter how convenient the interface.

My Bet: Mendral and similar tools will find their niche in exploratory analysis and hypothesis generation rather than production monitoring. The sweet spot is augmenting human intuition, not replacing it. Within 18 months, we'll see hybrid workflows where LLMs generate candidate queries that humans validate and refine—capturing most of the efficiency gains while avoiding the reliability pitfalls.