# Disarray

> Disarray builds AI research agents for long-horizon autonomy over heterogeneous data, turning complex proprietary data into production-quality machine learning models at a fraction of the time and cost of manual development.

Disarray was born out of UC Berkeley's RISELab and continues that lab's tradition of building systems that push the boundaries of data and AI. Disarray agents manage the end-to-end research loop: translating high-level goals, finding relevant data, forming hypotheses, writing and running code, interpreting partial or failed results, and deciding what to try next. The product tackles two challenges that hold back today's AI research agents: generating stronger hypotheses for efficient experimentation, and avoiding the failure modes that derail long-running agent loops.

A guiding insight: success comes from better context, not more context. Disarray's core differentiator is a context graph that captures rich lineage across data, code, execution history, and documentation, enabling precise retrieval of the most relevant information and multi-hop analysis that surfaces novel insights from prior work. With more targeted context, higher-order understanding, and intelligent task harnesses, Disarray's agents form better hypotheses, stay focused over time, and deliver breakthrough results rapidly in long-horizon research.

Every agent run produces a high-fidelity trace — what context was retrieved, what hypothesis was formed, what code ran, what errors occurred, how the agent recovered, and the final result. These traces are precise training data for improving every component of Disarray: today they self-heal the context graph, and dynamic self-improvement of the harness and planning modules is an active area of research. The instrumentation that runs research today is the foundation for recursive self-improvement tomorrow.

Principles guiding the product:

- Context is a core primitive, and quality matters more than quantity.
- Human experts stay in control of high-judgment decisions (which outcome definition to use, how to interpret missing data, what evaluation trade-offs to accept).
- Reuse prior work instead of starting from scratch; compound institutional knowledge over time.
- Fit into existing stacks (warehouses, feature stores, experiment trackers, orchestration tools) rather than disrupt established workflows.

ML model development is Disarray's first product use case, but the underlying capabilities extend to a much broader class of open-ended research problems. Disarray's agents have been validated on Kaggle: 28 medals across vision, NLP, tabular, and object detection competitions, including nine top-10 finishes and one better-than-human result. Each run was fully autonomous and limited to 24 hours on a single GPU.

Founded by Doris Xin (CEO; a decade of research and industry experience in production ML systems and model development automation; PhD at UC Berkeley's RISELab in collaboration with Google Research and Microsoft Research; early ML engineer at LinkedIn; first Databricks intern and Apache Spark MLlib contributor) and Moustafa AbdelBaky (CTO; two decades of autonomous-systems research spanning distributed systems at IBM Research as a three-time IBM PhD Fellow, edge-native ML as a UC Berkeley RISELab postdoc, and Technical Director at NASA Ames, where he received NASA's Early Career Achievement Medal).

Backed by 406 Ventures, Costanoa Ventures, Essence, Laude Ventures, and Zetta, with angel investors including Allie Miller, Andy Konwinski (Databricks and Perplexity co-founder), Anthony Goldbloom (Kaggle founder), DJ Patil (former U.S. Chief Data Scientist), and Rob Ferguson.

## Core pages

- [Home](https://disarray.ai/): Company overview, the end-to-end research loop, validation on autonomous ML experimentation, and the path toward recursive self-improvement.
- [Product](https://disarray.ai/product): Disarray's ML agents and the insights behind them — context as the bottleneck, the limits of full automation, wasted institutional knowledge, and the need to compose with existing infrastructure rather than disrupt it. Built on the context graph and a long-horizon agent harness, with humans kept in the loop for high-judgment decisions.
- [About Us](https://disarray.ai/about-us): Origins at UC Berkeley RISELab, founder backgrounds, and investors.
- [Featured](https://disarray.ai/featured): Index of press coverage and Disarray's own blog posts.

## Founder profiles

- Doris Xin, Co-founder and CEO: [LinkedIn](https://www.linkedin.com/in/doris-xin/), [Google Scholar](https://scholar.google.com/citations?hl=en&user=YuUE4doAAAAJ)
- Moustafa AbdelBaky, Co-founder and CTO: [LinkedIn](https://www.linkedin.com/in/moustafa-a/), [Google Scholar](https://scholar.google.com/citations?user=DlBr9YEAAAAJ)

## Blog posts

- [Agentic Search Is Not Enough: Why AI Agents Need a Context Graph](https://disarray.ai/featured/agentic-search-is-not-enough-why-ai-agents-need-a-context-graph): By Doris Xin, May 14, 2026. Argues that agentic search can retrieve fragments of context but cannot give agents a durable model of how an organization's systems relate, forcing them to rebuild cross-system structure at runtime — slow, brittle, and non-deterministic. A context graph supplies four things runtime search cannot: a data model of the organization, a unified index across scattered systems, a memory layer of prior decisions, and a computational substrate for structural queries (lineage, blast-radius, centrality, temporal "time travel"). Walks through construction (schema design, ingestion and versioning, entity resolution, relationship accuracy with calibrated intent-vs-execution and association-vs-dependency trust) and interaction (retrieval, analytical queries, and a feedback loop that lets the graph self-heal). Uses a running e-commerce example (Olist dataset) and two ML use cases to show where agentic search fails and a graph wins. Frames the context graph as the structural answer to enterprise "AI sprawl."
- [Design Principles for Long-Running Research Agents](https://disarray.ai/featured/design-principles-for-long-running-research-agents): By Moustafa AbdelBaky, April 28, 2026. Presents 11 design principles for the harness around a long-running research agent, organized into four areas: separation of concerns, durable state and curated memory, inspectability and observability, and agent-facing tool design. Core argument: research is sustained search through a hypothesis space, not task completion, and the hard problems (over-commitment to expensive trajectories, premature success claims, lost state across restarts, stale memory, reward hacking, silent multi-hour computations) are fixed by the harness, not a smarter inner loop. Principles include separating execution from evaluation, treating trajectory control and hard constraints as supervisory functions enforced outside agent control, building observability and durable state (anchored by the Context Graph) into the system, curating memory to preserve signal, and designing tools for machine consumption. Concludes that the harness should evolve as models improve, rebalancing responsibilities between model and system.
- [When an MLE Agent Beats Humans, What Does That Actually Mean?](https://disarray.ai/featured/when-an-mle-agent-beats-humans-what-does-that-actually-mean): By Doris Xin, April 3, 2026. Disarray's MLE agent earned 28 Kaggle medals across vision, NLP, tabular, and object detection competitions in fully autonomous runs capped at 24 hours and a single GPU each. Contrasts Disarray's autonomous-from-task-description approach with Karpathy's autoresearch (which requires hand-written program.md scaffolding). Walks through specific competitions — Google Research Contrails (gold), Cassava Leaf Disease (bronze in 13 hours), U.S. Patent Phrase to Phrase Matching (top 3%), and Tabular Playground May 2022 (2nd of 1,115 teams) — showing where the agent matched or diverged from winning human solutions. Argues that despite agent capability on the modeling task, human MLEs remain essential for production: business alignment, governance, accountability, and the broader ML lifecycle (feature stores, MLOps, monitoring, A/B testing, feedback loops). Goal is human augmentation, not replacement.

## Press

- [Before AI Can Train Itself, It Has to Learn What Its Data Means](https://aijourn.com/before-ai-can-train-itself-it-has-to-learn-what-its-data-means/): aijourn.com (The AI Journal), May 7, 2026. By Doris Xin. Argues the key barrier to automating model development is data understanding in proper context, not model capability. Frames the context problem as having three faces — fragmentation across systems, semantic ambiguity, and the absence of institutional memory — and explains why prior advances (AutoML, MLOps, coding agents) each solved real problems but left the context problem untouched. Connects enterprise ML automation to foundation-model training and the path to self-improving AI.
- [The Critical Path to Automating Model Development](https://www.unite.ai/autonomous-ml-engineering-data-understanding-bottleneck/): unite.ai, April 14, 2026. On the foundational challenges — particularly data understanding — that must be solved before ML model development can be meaningfully automated.

## Related research

- [Whither AutoML? Understanding the Role of Automation in Machine Learning Workflows](https://arxiv.org/abs/2101.04834): Co-founder Doris Xin's research paper grounding Disarray's stance on the necessity of human oversight for accountable AI systems.

## Contact and operations

- General inquiries: info@disarray.ai
- Careers: careers@disarray.ai
- LinkedIn: https://www.linkedin.com/company/disarray-labs/
- Security and trust: https://trust.disarray.ai/
- Platform status: https://status.disarray.ai/
- Privacy policy: https://disarray.ai/privacy
- Terms of service: https://disarray.ai/terms-of-service