Dr. Henry Kang, Founder and CEO of Clipto – Interview Series

Dr. Henry Kang, Founder and CEO of Clipto, is an entrepreneur and AI researcher with a track record of building successful AI-driven companies spanning computer vision, multimodal AI, and the creator economy. After earning a Ph.D. in Computer Vision from Carnegie Mellon University, where he conducted pioneering research alongside leading AI scientists and published influential work on large-scale object recognition, Kang held research roles at Microsoft Research and Intel before founding multiple startups. His previous ventures, Peekabuy and ZenVideo, were both acquired, with ZenVideo becoming part of Tencent, where he later served as Senior Director leading AI-generated content (AIGC), product management, user growth, and enterprise sales initiatives. In 2023, he launched Clipto, which reached profitability within three months, surpassed $10 million in annual recurring revenue during its first year, and now serves millions of users in more than 150 countries, reflecting his continued focus on bringing advanced AI technologies into practical, high-impact applications.

Clipto is a Palo Alto-based AI company developing an on-device multimodal content platform that enables professionals to search, organize, transcribe, and analyze video, audio, and image libraries without uploading data to the cloud. The company’s AI models run locally on users’ devices, emphasizing privacy, low latency, and offline functionality while reducing cloud infrastructure costs. Designed for creators, researchers, legal professionals, healthcare organizations, and enterprise teams, Clipto transforms large multimedia collections into searchable knowledge bases through natural language queries, automatic transcription, speaker recognition, and intelligent content indexing. Since its founding in 2023, the company has attracted millions of users worldwide, secured backing from prominent investors including HSG, GL Ventures, EnvisionX Capital, Palm Drive Capital, and Hans Tung, and is building what it describes as an AI-native operating system for multimodal content workflows.

What first convinced you that the next major opportunity was not just generating more content, but helping people understand, search, and reuse the media they already have?

Over the past three years, generative AI has dramatically reduced the cost of creating information. But creating more information has also exposed a much bigger problem: almost none of it is truly reusable.

We don’t have a creation problem anymore. We have a memory problem.

What convinced me wasn’t a technical breakthrough. It was watching professionals repeatedly fail to find information they already owned. Editors couldn’t locate a quote from an interview they recorded two years ago. Researchers reread reports they had already written. Marketing teams recreated assets because they couldn’t rediscover them.

People remember ideas, conversations, moments, and insights. Computers still organize information through filenames, folders, and metadata. That mismatch is where enormous amounts of knowledge become effectively lost.

I believe the next opportunity in AI is not generating more content, but making existing knowledge continuously accessible and actionable. If AI can instantly recover a forgotten insight from hundreds of hours of footage, reconnect related information across years of work, or surface exactly the right moment from an archive, it fundamentally changes how people think and work.

Search is simply the interface. The real opportunity is building the persistent memory system AI has always been missing: one that continuously understands, organizes, and preserves everything people already know, making that knowledge available to both people and the AI agents that increasingly work alongside them.

How should companies think about measuring AI ROI when the main benefit is removing hours of searching, organizing, and manual review?

AI ROI is often measured through automation rates, cost reduction, or headcount efficiency. Those metrics matter, but they overlook one of the largest hidden costs inside knowledge work.

Search is invisible work. Nobody budgets for it, yet everyone pays for it.

An editor spending two hours locating the right clip, a researcher reviewing days of interviews, or a marketing team rebuilding assets that already exist are all paying what I call a search tax.

AI creates value by reducing search and review cycles from hours to seconds.

The impact extends well beyond time savings. It increases throughput, shortens production cycles, improves the utilization of existing content, and allows professionals to spend more energy on creative and strategic work.

Ultimately, AI ROI should not be measured only by what organizations spend less on, but by what they become capable of producing. Time returned becomes additional creative capacity, faster iteration, better decisions, and ultimately better outcomes.

Why has media search remained such a persistent bottleneck, even as content creation tools have become more advanced?

Content creation has evolved dramatically over the past decade, but the way we organize media has changed surprisingly little. Most systems still rely on folders, filenames, timestamps, and manual tags.

The problem is simple:

Humans remember meaning.

Computers remember filenames.

A creator rarely thinks, “What was the file name?” Instead, they ask, “Find the interview where the founder talked about pricing,” or “Show me the scene where someone walked into a coffee shop wearing a red jacket.”

Traditional media management systems were never designed to answer questions at that semantic level.

As personal and professional media archives continue to grow, the gap between how humans remember information and how computers store it becomes increasingly costly. Media search remains a bottleneck because our tools are still optimized for managing files, while people are trying to retrieve meaning.

What does AI need to do differently to become useful without forcing people to reorganize how they already work?

Creative professionals have spent years refining workflows that match the way they think and work. Asking them to reorganize folders, rename files, rebuild asset libraries, or adopt entirely new processes inevitably creates friction.

I believe AI should adapt to people—not force people to adapt to software.

The best AI systems integrate naturally into existing workflows. They work quietly in the background, understand existing content automatically, and let users interact through natural language instead of rigid organizational rules.

People should never have to decide in advance where information belongs or how it should be tagged. They should simply ask for what they remember.

Great AI should disappear into the workflow. When users stop thinking about the software and simply trust that the right knowledge will always be there when they need it, AI has become infrastructure rather than another application.

How do you define the “search tax” in creative work?

Search tax is the hidden cost of information you already own.

More formally, it’s the cumulative time and cognitive effort people spend trying to locate knowledge they already possess.

It is one of the least visible, yet most pervasive, inefficiencies in creative work.

Editors revisit hundreds of clips to find a five-second sequence. Journalists review hours of interviews to recover a quote. Marketing teams recreate assets because existing work cannot be easily rediscovered.

Each individual moment seems insignificant. Together they consume enormous amounts of time and attention.

Reducing search tax doesn’t simply save time. It unlocks the full value of knowledge organizations have already created.

Why is local, on-device AI important for professionals working with sensitive content?

Many professionals work with material that simply cannot be uploaded to external systems.

That includes unreleased productions, client footage, legal recordings, confidential interviews, proprietary research, and personal archives accumulated over many years.

Traditional MAM platforms were built around metadata management, while many cloud-based AI systems require content to leave the user’s environment for analysis. Clipto takes a different approach by bringing multimodal intelligence directly to the data.

Local AI processing improves privacy, reduces latency, lowers compliance concerns, and gives users stronger control over governance and access.

We believe privacy-first AI is moving from being a competitive differentiator to becoming a professional baseline.

Privacy should be an architectural property,

not a policy.

Why will memory and retrieval become one of the most important layers in the AI stack?

Memory and retrieval will become one of the most important layers in the AI stack because they make AI personal, contextual, and continuously useful.

Foundation models understand the world. Memory systems understand your world. Agents connect the two. But neither inherently understands an individual’s long-term context—meetings, conversations, projects, creative assets, decisions, or accumulated experience.

Without memory, every interaction starts over.

That is the gap Clipto is solving. We are building a persistent memory layer on top of users’ multimodal data—video, audio, images, documents, notes, and transcripts—turning fragmented information into structured, searchable, reusable knowledge that both people and AI agents can continuously build upon.

Over time, the most valuable AI systems will not be defined solely by the foundation models they use, but by the quality of the memory they can access and the relevance of the context they retrieve.

Just as databases became the persistence layer for software, memory will become the persistence layer for AI.

What are the technical challenges behind multimodal search?

People expect search to feel effortless. Making that happen is technically difficult.

Multimodal search is not simply retrieval over embeddings. It is a complete media understanding system that transforms unstructured content into searchable memory while balancing accuracy, latency, cost, privacy, and scalability.

The system must understand speech, text, visual content, people, objects, actions, temporal relationships, and semantic intent simultaneously. It must also deal with ambiguity because humans rarely remember exact filenames or timestamps.

Someone might ask, “Find the scene where fundraising was discussed outside a conference center.” Answering that requires language understanding, visual reasoning, temporal alignment, cross-modal fusion, and effective ranking.

In practice, retrieval quality depends far more on representation, indexing,  ranking and systems design than simply embedding everything into a vector database.

Making multimodal retrieval feel natural is ultimately a systems problem, not a single-model problem.

Where does Clipto fit into the shift from AI generation toward knowledge infrastructure?

The first wave of AI was about generation.

The next is about understanding.

The one after that is about memory.

And memory is what makes AI agents continuously useful.

Clipto is building that missing layer.

Our goal is not simply to help people generate more content, but to continuously organize, preserve, and activate everything they already know.

We see search as the interface, not the destination. The real product is persistent memory infrastructure—a persistent layer that gives both people and AI agents access to the right context at the right moment.

What will separate valuable AI tools from software clutter?

The most successful AI tools will not be the ones with the biggest models or the longest feature lists.

They will be the ones people eventually stop noticing.

People do not adopt software because it is intelligent. They adopt it because it removes friction, reduces complexity, and consistently helps them accomplish meaningful work.

AI that asks users to change behavior, duplicate effort, or manage yet another application will eventually become noise.

The best AI products won’t feel like AI products.

They’ll simply become part of how people think, remember, decide, and create.