Hi friends! Welcome back to another installment of our newsletter, we have a lot to fill you in on.

Most enterprise data is unstructured. Code, documents, emails, Slack threads, meeting notes: the information that captures how an organization thinks and operates largely lives outside the clean rows and columns of a database. Estimates put it north of 80% of all data generated yearly, and until recently, software systems mostly ignored it because it was too hard to query reliably.

AI changes that equation. Unstructured data is suddenly the most valuable signal in any company's stack, but unlocking it requires infrastructure that doesn't yet exist at production quality. The indexing problem is hard because data is heterogeneous, constantly changing, and the semantics that make it useful are fragile. Pipelines that work well for PDFs don't work well for code; pipelines that handle Slack messages well tend to fall apart on legal contracts.

This is the space CocoIndex is building into.

The framework is open-source, Rust-backed, and built around a declarative, incremental model. You define what your index should look like and the engine keeps it in sync with reality, reprocessing only what changed. Think of it as React for data engineering: write your transformation logic once, and the framework handles the delta forever. The result is AI systems that stay current without the compute overhead of full re-indexing and without the stale context that makes today's RAG pipelines frustrating in production.

A concrete example: meeting notes synced from Google Drive, incrementally extracted into a live knowledge graph so that an agent querying "what did we decide about the pricing model last quarter" gets the answer as of this morning, not last Tuesday's batch run. An agent with stale context isn't helpful, and in many enterprise settings it's actively dangerous.

What's strategically interesting about CocoIndex is the data-type-by-data-type playbook. Rather than trying to build a generic unstructured data layer all at once, each data type gets its own purpose-built application on top of the core engine. The framework handles the hard infrastructure: incremental processing, lineage tracking, state management. The application layer handles the domain-specific semantics that make retrieval actually useful.

CocoIndex-code is the clearest example. Modern coding agents (Claude Code, Cursor, Codex) are powerful but structurally blind to global codebase context. They miss the function that already exists three directories over, which is something every engineering team is quietly experiencing. CocoIndex-code plugs this gap by parsing codebases via Abstract Syntax Trees rather than line count, running as a lightweight MCP server, and cutting token usage by 70% per turn by surfacing only what's contextually relevant.

But fresh context is only half the problem. The other half is what happens to the context agents produce.

This is where Mesa becomes interesting. As agents increasingly work on long-lived documents, including contracts, audit reports, case files, and codebases, the artifacts they generate need to be versioned, shared, and collaboratively refined by other agents and by humans. Mesa is a POSIX-compatible, durable filesystem with built-in version control designed specifically for this use case. Agents can work in parallel on isolated branches without locking, every change is automatically checkpointed and rollback-able, and per-agent access controls with full audit history make it viable in regulated environments. The mental model is Git, but for everything an agent touches, not just code.

Together, CocoIndex and Mesa gesture at a more complete vision of the emerging stack: infrastructure that keeps input context perpetually fresh on one end, and infrastructure that makes agent-produced context durable, collaborative, and trustworthy on the other. The GTM logic for both is the same: start with code and developers, prove the thesis, then expand the playbook across every unstructured data type that matters.

As agents become the primary interface to organizational knowledge, the infrastructure layer underneath them becomes critical path. Unstructured data is where the value is, and the teams building the picks and shovels to make it usable are worth watching closely.

Keycard, Herald, Restate, DragonflyDB, and Modal were all named to the Redpoint 2026 InfraRed 100, which recognizes the top private companies shaping the future of cloud infrastructure. Congrats to our founders 👏

Modal just raised a $355M series C, valuing it at $4.65B. They've grown 500% in 8 months by betting that the cloud built for traditional web applications was never going to fit AI workloads. Modal has now become one of the fastest-growing compute platforms for AI training and inference 🚀 Read more here.

Fearn is building the infrastructure layer for IP law, and they just announced their $5.5M seed round 🎉 The team has built an entirely self-hosted product composed of dozens of AI models that compresses 50+ hours of work into 20 minutes. Read more here.

RunLLM is now Herald. The new name reflects the fact that their AI SRE is the only product on the market that operates autonomously. It teaches itself about your product & infra, detects early warning signs of incidents, and investigates without runbooks. Read more here, or try it out!

MagicPatterns is an AI prototyping tool for product teams that turns prompts into production-ready UI with design system matching and Figma/code export. Check out the evolution of their design agent here, or try it out!

Here’s our most recent podcast episodes from The Infra Pod with Shubo, CTO of Axiom Math, and the Open Source Startups Podcast with James Bayer, Chief Product Officer at Flox.

Restate continues to earn strong customer praise even as they scale quickly.

They are currently hiring across engineering, design, and GTM. See open roles here.

Many of our other portfolio companies are hiring across similar roles. Explore more here!

Tim Chen is speaking at the Databricks Data + AI Summit Startup Forum on June 16th in San Francisco alongside VCs from a16z and NEA and founders building the next generation of AI-native companies. If you're attending the summit, come say hi!

Essence VC is co-hosting another GTMfor.dev live podcast on Wednesday June 17th, from 1-2pm PST. We will be interviewing Zach Lloyd, the CEO of Warp, on how he took on one of the hardest GTM challenges in developer tools: getting millions of developers to switch away from the terminal they've used their entire careers. We will be going over how to win on switching costs, how to get your first 1000 true believers, monetization, open-source, and more. We’d love to see you there!

CocoIndex co-founder Linghua Jin is speaking at Harness Engineering: State of the Art in Agent Harnesses on June 29th at 6pm at the AWS Builder Loft in San Francisco. Learn how coding agent harnesses are designed, what breaks in real-world deployments, and how teams are building more robust agent architectures. Save your seat here.

Thanks for reading! We'll be back next month with more updates.

— The Essence Team