Empirik Raises $21M—Its AI Tries to Stop Outages Before Deployment

Empirik emerged from Sequoia Capital’s incubation program as an independent company on September 1, 2026, with $21 million in seed funding. TechCrunch’s September 1 coverage describes a system that tracks proposed infrastructure changes, infers their effects across connected systems and flags the most dangerous updates for human review.

The startup is trying to move a reliability decision ahead of deployment: estimate what a change could affect while engineers can still block or modify it, instead of waiting for production telemetry to reveal a failure. Enterprise use has been disclosed, but no public benchmark yet demonstrates that the product prevents outages consistently or can execute broadly without human supervision.

Funding brings an enterprise product out of stealth

Empirik’s launch post states that it secured more than $21 million from Sequoia, S32, Canapi Ventures and Alumni Ventures; identifies production environments at Guardant Health, an unnamed Fortune 50 consumer-packaged-goods company and an unnamed Fortune 500 financial-services company; and places autonomous execution later in a progression that begins with answering infrastructure questions.

The distinction between “$21 million” and “more than $21 million” reflects the sources’ wording rather than evidence of separate rounds. The disclosed deployments establish that organizations are using the product in production, but they do not reveal contract values, the number of changes evaluated, infrastructure coverage or how much execution authority any customer has granted it.

Empirik grew from work started inside Sequoia in 2023 by infrastructure executives Avon Puri and Sudheer Dhurjati. Kartik Chandrayana, previously chief product officer at Quantum Metric and an observability executive at Salesforce, became chief executive as the project moved toward an independent company.

From a declared change to a human decision

Empirik’s proposed workflow begins with the change rather than an alert. Its product is intended to combine the action’s declared intent with a current model of the customer’s cloud, software-as-a-service and on-premises environment, then determine which connected resources and owners could be affected.

1. Capture the proposed action. The system receives the intent of an infrastructure change before execution.
2. Compile the dependency graph. It models relationships among resources, including virtual constructs such as identity policies and routing tables that may not appear in runtime traffic maps.
3. Calculate the blast radius. The system follows relevant relationships to identify potentially affected infrastructure and responsible owners.
4. Apply the policy gate. A lower-risk action may proceed, while a larger one can receive conditions or guardrails.
5. Escalate material risk. Changes judged most dangerous remain subject to human review before deployment.

This sequence does not eliminate the need for observability. Logs, metrics and traces still show how running systems behave and help diagnose incidents. Empirik is addressing an earlier question: what could happen if a particular change is allowed to execute?

Where dependency mapping ends and predictive AI begins

The deterministic part is clearest when a dependency is already represented in the graph. If a recorded routing relationship connects a network to several workloads, graph traversal can reproducibly identify those workloads as potentially affected. A fixed rule requiring approval whenever a protected production resource appears in that result can also produce a deterministic decision.

The reliability of that answer still depends on the graph being complete and current. Infrastructure-as-code describes intended state but may omit live relationships; runtime topology observes communication but can miss constructs such as IAM policies, quotas and security groups; short-lived resources can also outpace conventional inventories. A deterministic calculation over an incomplete model can return a precise but incomplete blast radius.

Prediction starts when the product must infer a relationship that has not been directly recorded, interpret ambiguous intent, estimate whether an affected component will actually fail or rank several possible consequences. Empirik’s public material does not specify which graph edges are observed and which are inferred, what models assign risk, or how uncertainty is shown to reviewers. Its claim of deterministic impact calculation therefore describes the intended result, not public proof that every relevant dependency is known.

Production use is not yet evidence of outage prevention

The available evidence supports a narrower conclusion than the product’s ambition. Empirik has financing, named and unnamed enterprise deployments, and a pre-deployment change-analysis model. It has not released a controlled comparison connecting its recommendations to fewer incidents.

Tech Times’ launch review notes that change-impact accuracy, false-positive rates, coverage gaps in partially instrumented environments, the underlying model architecture and production error rates have not been disclosed. Without those measures, outsiders cannot determine how often the system catches a dangerous change, blocks a safe one or misses an important dependency.

The autonomy boundary is also material. The published roadmap moves from answering questions to validating actions and, eventually, executing them as customers grant more authority; meanwhile, the described workflow retains human review for the highest-risk changes. Current production use should therefore not be read as evidence that customers have handed Empirik unrestricted control of their infrastructure.

The next test is whether Empirik or its customers release comparable operating data: prevented incidents, missed dependencies, false alarms and the share of changes allowed to execute automatically. Until then, the company’s strongest verified claim is that it can place dependency analysis and risk review before deployment—not that it has already proved outages can be reliably prevented.

Also read:

• Why America Has Forgotten How to Build While China Can't Stop: The Great Infrastructure Divergence
• Startup Option Pools: Calculate the Pre-Money Dilution Before You Sign
• TerraFirma Raises $115 Million for Infrastructure Construction Technology