What is Cascade?

Cascade builds custom evaluation infrastructure that makes AI agents reliable by learning from their real production behavior. Part of YC W2026, it was founded by Adam AlSayyad (CEO) and Haluk Cem Demirhan (CTO), both researchers from the Berkeley AI Research (BAIR) Lab — the same lab behind Databricks and Perplexity.

Most deployed AI agents remain static after launch, with teams manually adjusting prompts and inspecting logs without reliable ways to measure alignment. Cascade solves this by observing real production runs, training evaluator models that learn what "correct" looks like for a company's specific workflows, and converting those judgments into training signal for continuous improvement.

The platform targets the gap between general-purpose LLMs and enterprise-specific operational needs, helping organizations develop specialized models aligned to their unique data and processes. Cascade is already deployed in legal reasoning workflows and high-volume customer support. The AI guardrails market they target is projected to grow from /bin/zsh.7B (2024) to B by 2034.

Key features

Core capabilities this platform advertises.

Model distillation
Proprietary-to-small model conversion
Cost reduction
Portable intelligence

Strengths and tradeoffs

What this tool does well, and the limitations to keep in mind.

Pros

Elite research pedigree from BAIR (Berkeley AI Research) lab
Already deployed in production for legal reasoning and customer support
Addresses real pain point of agents degrading silently post-deployment
Custom evaluation models learn company-specific definitions of correct behavior
CTO has production-scale experience from Netflix and Amazon

Cons

Plans & pricing

What's included in each plan, and how the tiers compare.

Custom

Contact for pricing

Custom evaluation models
Production behavior learning
Continuous improvement pipeline
Enterprise deployment

View official pricing page

Common use cases

Teams wanting proprietary model quality at lower cost

Model compression
Cost optimization
On-premise deployment of distilled models
Edge deployment

Using Cascade with Respan

Cascade evaluates and improves AI agent behavior over time, while Respan provides real-time monitoring of the LLM calls driving those agents. Together they enable both continuous improvement and real-time observability.

Feed Respan production traces into Cascade for evaluation model training
Monitor LLM performance with Respan while Cascade evaluates output quality
Use Cascade-trained evaluators alongside Respan monitoring for comprehensive agent oversight