Krishna
Gade

The Audit Trail

In 2017, Facebook's News Feed was under fire. Billions of posts, ranked by algorithms nobody could see, shaping what a billion people believed. Krishna Gade was the engineering manager whose team got handed the hardest question in tech: how do you explain a machine's decision to a human being?

The result was "Why am I seeing this?" - a small button that opened a window into the black box of algorithmic recommendation. For the first time, at scale, users could ask a platform why it thought they needed to see a particular post. Gade's team had built the industry's first large-scale AI explainability feature, deployed to billions.

He spent a year sitting with what that meant. Teams at every company were deploying AI with no idea what their models were actually doing. The decisions those models made - who gets a loan, which resume gets flagged, what news travels - had real consequences. And nobody was watching.

In October 2018, Gade left Facebook and co-founded Fiddler Labs - alongside Amit Paka and Manoj Cheenath - in Palo Alto. The original pitch was almost humble: a "Tableau-like tool for machine learning," something that would make model behavior as visible as a bar chart. Lightspeed Venture Partners and Bloomberg Beta wrote the first check in a $3 million seed round. They had seen something.

What they had seen was a category waiting to be named. Fiddler called it AI Observability. Model monitoring, explainability, bias detection, fairness assessment - assembled into a platform that sits between an enterprise's AI and the consequences of its decisions. The US Navy signed on. Fortune 500 banks and insurance companies followed. The CIA's venture arm, In-Q-Tel, invested.

By January 2026, Fiddler had raised $30 million in a Series C led by RPS Ventures, bringing total funding past $112 million. The question Gade's team answered at Facebook had become a $100-million-plus enterprise business - and with AI agents now making decisions that no single human ever reviews, he argues the problem is only getting bigger.

Infrastructure First

Gade came to the US from India to study computer science at the University of Minnesota, finishing his MS in 2004 with research focused on document clustering - a quiet foreshadowing of a career spent wrangling information at scale.

He joined Microsoft, where he spent five years on the Bing Search Engine - building index quality measurement systems, static rank infrastructure, web spam detection, and query suggestions. The web in 2004 was still mostly a document retrieval problem. Gade was solving it at Microsoft's scale.

Twitter came next. In the early 2010s, when social media was transforming what "real-time" meant for information systems, Gade led the web search and real-time search engineering teams. He built streaming data infrastructure using Apache Storm - one of the early production deployments of stream processing at that scale. His systems tracked the news as it happened.

Pinterest was where Gade learned to build teams as much as systems. He joined as a two-person data engineering function and grew it to 25+ engineers. He built the A/B testing framework, the scalable logging pipeline, fast interactive query infrastructure, and self-service batch processing that handled hundreds of petabytes. Pinterest's data team became a model for how data infrastructure gets done at a fast-growing consumer platform.

Then Facebook, and the News Feed ranking problem. Gade ran the engineering team responsible for understanding why the most-read content feed in history surfaced what it surfaced. "Why am I seeing this?" was not a regulatory box to check. It was, for Gade's team, the answer to a genuine engineering question: can you make the algorithm legible?

The answer was yes - but it required a new kind of infrastructure. Infrastructure for observability. Infrastructure for accountability. That was Fiddler.

Fiddler AI: The Trust Layer

Fiddler AI is not in the business of building AI models. It is in the business of watching them. The platform sits in production between a company's AI system and the decisions it produces - monitoring for model drift, detecting bias, surfacing explainability, tracking data distribution shifts, and generating audit trails for regulators.

The original use case was predictive models in finance and insurance - high-stakes decisions where a wrong prediction meant a loan denial or a claims refusal, and where regulators increasingly wanted to know why. Fiddler gave risk teams the ability to see which features drove a model's output, when those features started shifting, and where fairness metrics were degrading.

The platform evolved as the AI landscape did. When large language models moved from research to production, Fiddler expanded into LLM observability - monitoring hallucination rates, input/output quality, guardrails for sensitive content, and latency across multi-model chains. The product today covers the full lifecycle: from predictive models to generative AI to agentic systems.

In 2025, Fiddler acquired Lumeus.ai to extend its reach into AI agents - systems that autonomously browse, execute code, call APIs, and make sequential decisions. Gade has been direct about the concern: "There's a growing gap where organisations lack control over the behaviour of these agents" - where agents produce insecure code, expose sensitive data, misuse tools, or execute unintended workflows at scale. The acquisition gives Fiddler observability at both the creation layer and the runtime layer of agentic AI systems.

Algorithmic Justice

Gade has a phrase he returns to: "Algorithmic Justice." It is not marketing copy. It is his argument that every system making decisions that affect humans - hiring, lending, healthcare, content - has an obligation to be auditable, and that building the tools to make that auditing possible is itself a form of justice work.

The argument is grounded in specifics. Bias in lending models has measurable disparate impact on protected classes. Data drift in healthcare models means a classifier trained on one patient population quietly degrades when deployed to another. Hallucinating LLMs in customer service applications put companies at legal risk while confusing real people. These are not hypotheticals - they are production incidents that Fiddler's customers are trying to prevent.

The platform reflects the philosophy. Fiddler does not just alert when a model underperforms - it provides root cause analysis. It does not just flag potential bias - it quantifies fairness metrics across demographic attributes and explains which input features are driving the disparity. It does not just monitor LLM outputs - it offers guardrails that intercept unsafe or off-topic responses before they reach users.

The "four pillars" Gade describes for model performance management - monitoring, root cause analytics, explainability, and fairness - are not a product brochure. They are a framework for what responsible deployment actually requires. Building a model is no longer enough. Watching it is the job that comes after.

Milestones

What He Built

• Led creation of "Why am I seeing this?" at Facebook - the first high-scale AI explainability feature deployed to billions of users
• Co-founded Fiddler AI, raising over $112 million across five funding rounds from Seed through Series C
• Attracted strategic investment from Lockheed Martin, Amazon Alexa Fund, In-Q-Tel (CIA), Mozilla Ventures, and Cisco Investments
• Scaled Pinterest's data engineering team from 2 to 25+ engineers while building petabyte-scale infrastructure
• Built Twitter's real-time search and streaming data infrastructure at one of the earliest production deployments of Apache Storm
• First company to offer a unified AI Control Plane at both agent creation layer and runtime layer
• Fiddler AI now serves Fortune 500 enterprises and US government agencies including the US Navy
• Pioneer of the AI observability category before the category had a name
• Speaker at Amazon re:MARS, NYSE Floor Talk, Montgomery Summit, MachineCon, TWIMLcon, and O'Reilly Strata

On Video

Gade speaks with infrastructure clarity and startup urgency. These conversations span his views on AI accountability, model governance, and the emerging challenge of agentic AI systems.

The Vocabulary He Works In