[Interview] Meet HEIMDEX: The AI SaaS Redefining Video Search, Summarization, and Editing Without the Uploads

Published at Jun 18, 2026

- An Interview with Jangwon Lee (CEO) and Heejo Kang (CSO), Founders of Antler Korea’s Portfolio Team, Heimdex
- Natural Language Search Without the Uploads: Re-architecting AI Utilization in Video
- Expanding from entertainment to live commerce, law firms, investigations, and research.

Video has undeniably become the foundational data format of both the internet and modern enterprise. Companies and organizations generate and store hundreds of hours of footage daily. The real challenge, however, begins after the camera stops rolling. Storing video does not automatically turn it into an accessible asset. In fact, as video data grows, finding a specific scene becomes increasingly difficult.

For years, the video industry has harbored a frustrating paradox: while data is abundant, retrieving it still relies entirely on human memory and manual labor. In broadcasting and entertainment, editors routinely reopen dozens—sometimes hundreds—of files just to find a single, specific cut.

This bottleneck is not exclusive to show business. In the legal sector, lawyers manually cross-reference case-related videos, photos, and recordings to stitch together a timeline. In sensitive fields like criminal justice, security, insurance adjustment, and corporate research, the problem intensifies. Because this data is highly confidential, organizations cannot simply upload it to cloud-based AI due to strict privacy regulations. As a result, operations stall at the most fundamental stages: search and classification.

Heimdex offers a breakthrough solution to this age-old problem, drastically cutting down both time and costs. They have introduced an AI SaaS platform capable of searching, summarizing, and editing videos using natural language—all without requiring an upload.

At the core of their platform is what they call the "Vector-Native Technology-based Hybrid Video Understanding Engine." By leaving the original video files exactly where they are on the user's local storage, the engine generates lightweight "sidecar metadata" composed of scenes, dialogue, OCR, subtitles, and embeddings to drive search and workflows. The goal isn't to make video AI more "flashy," but rather to transform the massive, dormant video archives locked inside organizations into highly functional assets.

"Sifting Through 100 Hours for a 10-Second Clip" — A Problem Rooted in the Field, Not the Lab

CEO Jangwon Lee is a software engineer from the University of Illinois Urbana-Champaign (UIUC). With a career spanning NHN, DeepSales, and Esri, Lee brings a rare blend of deep technical development and business acumen to the table. Co-founder and CSO Heejo Kang previously managed AI product management and B2B strategy at LG Uplus, where he witnessed the friction points of the video and media industries firsthand. The two met during the 7th cohort of the Antler Korea Generator program, where they quickly realized their shared conviction that video data bottlenecks could be solved through engineering.

Heimdex did not start with a grandiose technological vision or macro-market analysis. Instead, the spark came from a very specific, gritty industry reality. Reflecting on his early days as a producer at SM Entertainment, CSO Heejo Kang noted that his first assignment perfectly encapsulated the structural issues of the industry.

"I started my career as a PD at SM Entertainment," Kang recalled. "As the rookie on the team, my very first job was responding to seniors asking, 'Hey, find me this specific clip.' I had to dig through external hard drives containing over a hundred episodes, find the exact cut, and bring it over to their editing timeline. That was eight years ago, but the exact same workflow persists today. I knew that if we applied AI correctly, we could completely eliminate this incredibly labor-intensive bottleneck."

As Kang pointed out, modern video production environments still rely on outdated workflows. The more video an organization accumulates, the more convoluted the management system becomes, leaving the "search" process highly inefficient. Editors have to track down whoever remembers the clip, or guess random filenames. Once that person leaves the company, that institutional memory is lost forever. Ultimately, companies find themselves sitting on mountains of data that require prohibitive amounts of time and money to actually use.

While major broadcasters use massive Media Asset Management (MAM) systems, these platforms require human logging teams to manually watch footage and annotate metadata—such as cast members, locations, episode numbers, dialogue, and context—field by field. They also incur staggering cloud storage costs.

"For smaller organizations, traditional MAM systems are financially impossible," CEO Jangwon Lee explained.

"Traditional MAM systems require you to rent massive servers, upload all your video files, and hire people to manually tag everything. We realized AI could automate this entire process. Our goal from day one was to strip away the heavy, complex infrastructure of traditional asset management, lighten the architecture, and deliver a seamless, frictionless editing experience where users can find what they need instantly."

Furthermore, Kang highlighted that recent cloud-based video AI solutions still miss the mark for many teams. While technically impressive, they introduce new bottlenecks: the massive cost, time, and security risks associated with uploading raw, heavy video files to third-party clouds.

"Enterprise MAM deployment used to start at around $1 million. About four years ago, AI models emerged that allowed cloud-based searching. But smaller and mid-sized production companies couldn't afford the exorbitant cloud egress fees and processing times. We built a system that elegantly solves both problems simultaneously."

Heimdex’s addressable market extends far beyond media and entertainment. The founders see immense potential in public legal systems and corporate operations. For instance, law firms must manually review endless hours of evidentiary footage, and insurance adjusters spend hours pinpointing the exact second an incident occurs. While the terminology and final deliverables vary by industry, the pain point is identical: finding the exact right frame within a sea of video and image data.

During development, Kang leveraged his extensive entertainment industry network to gather feedback from active professionals, confirming a strong willingness to pay for a working solution. Along the way, they secured early-stage customers ready to adopt their Minimum Viable Product (MVP), validating their Product-Market Fit (PMF).

"We targeted the entertainment industry first because of our network, but also because its intellectual property (IP) framework is highly structured," Kang said. "Tracking a fixed cast of characters made it an ideal environment to rapidly validate our MVP. After securing funding, we split our product pipeline: Product 1 tailored for entertainment, and Product 2 built specifically for law firms, where lawyers must meticulously review every piece of video and photographic evidence for litigation."

In short, Heimdex isn't commercializing "video generation." Instead, they are unlocking the latent value of existing, proprietary data that companies have already spent money to shoot and store. Rather than chasing flashy generative AI demos, Heimdex targets the unsexy but critical bottleneck of enterprise data utility.

"Search Without Uploads" — A Vector-Native Architecture That Leaves Raw Video Untouched

Aside from cost and time efficiency, Heimdex's primary value proposition is its ability to search, edit, and utilize video via natural language without uploading files.

Their Vector-Native architecture processes original video files directly on local workstations, deconstructing them into scenes, audio, and text. The engine then vectorizes these components into highly compressed "sidecar" metadata files. The system queries this metadata index to execute search APIs and map directly back to the local editing timelines.

Think of it as leaving the heavy books on the shelf while the AI builds a highly structured, instantly searchable index system. This allows users to simultaneously search across fragmented storage environments—whether the files live on local drives, cloud storage, HDDs, or local NAS networks. Lee contrasted this with traditional MAM systems:

"Traditional media asset management requires you to ingest and process the entire video file, meaning you have to upload it somewhere and tag it manually. With Heimdex, the raw video stays exactly where it is. Our AI watches it once locally, logs the conceptual data, and syncs only that lightweight metadata to the cloud. By moving data that is fractions of the size of raw video, our solution deploys into active production environments exponentially faster."

Standard cloud-based video AIs typically charge per hour of video processing on top of ongoing cloud storage fees, often introducing massive upload latencies and breaking integration with local editing software. Heimdex circumvents this by offering a hybrid approach.

"The majority of video AI companies actually generate their margins from cloud storage and indexing fees rather than the AI models themselves," Kang noted. "They can't afford to abandon that revenue stream. When you combine cloud storage with heavy AI model indexing, it becomes cost-prohibitive for most teams. Because our architecture minimizes cloud storage to near zero, we possess a permanent cost advantage over the competition."

Equally impressive is how Heimdex redefines "search" from a strict keyword-matching exercise into a contextual, semantic experience. Traditional systems require exact matches; misspell a tag or a name by one character, and the system returns zero results. Heimdex vectorizes scenes, dialogue, faces, and text (OCR), allowing users to search based on abstract concepts, descriptions, or intent.

"In older systems, if your query was slightly off, you found nothing," Lee explained. "For example, if you misspelled the name of this interview location, traditional search would fail. But because our engine maps everything semantically via vector embeddings, you can use completely different phrasing or approximate descriptions and still pull up the exact scene you're visualizing in your head."

Furthermore, Heimdex customizes its indexing engine to learn industry-specific terminology. Entertainment networks, law firms, commercial production houses, and police departments all use radically different jargon and workflows. Heimdex believes the true moat in enterprise AI lies not in the base model, but in deep integration with domain-specific workflows.

Focusing on Utility Over Generation: Expanding into Legal, Research, and Investigation

Throughout the interview, the founders repeatedly emphasized that they are positioned at the exact opposite end of the spectrum from "generative video AI." While the tech sector rushes toward automated video generation and AI-synthesized content, Heimdex serves clients whose data cannot—and should not—be fabricated.

"We intentionally steer clear of the generative space because our clients deal with irreplaceable, real-world footage," Lee stated. "You aren't going to replace an A-list actor you paid millions of dollars to shoot with a generative AI substitute. And in fields like CCTV security or legal evidence, generating data is an outright liability. The market desperately needs tools that maximize the utility of existing, authentic footage. AI doesn't always have to mean 'generating' something from scratch."

Kang connected this philosophy to the broader enterprise issue of "dark data"—vast stores of unindexed, unusable corporate information. Heimdex’s long-term vision is to structure the estimated 90% of global video data that currently sits dormant, transforming it into highly searchable assets that can eventually feed back into the wider AI ecosystem. Even generative models, they note, fundamentally require structured, highly accurate training data. Heimdex aims to be the robust infrastructural bedrock beneath the video AI boom.

"Look at it this way: even generative AI video models require massive datasets to train on," Kang added. "One of our ultimate milestones is to serve as the bridge that structures this 90% of 'dark data,' making it available for model training or providing highly organized raw footage repositories. That’s a massive differentiator for us."

This thesis underpins Heimdex's aggressive market expansion strategy. While they launched within entertainment, their pipeline now includes law firms, insurance agencies, criminal investigation units, robotics labs, academic research institutions, live commerce brands, and independent production agencies. The common denominator? High video volume, high urgency, and strict security requirements.

"Initially, we laser-focused on just law firms and entertainment," Lee shared. "Live commerce wasn't even on our radar. But after hearing about our tech, a major live commerce enterprise reached out explaining they faced the exact same bottleneck. We are now co-developing a product tailored directly to their workflow."

The pain points vary distinctly by vertical. For entertainment clients, speed is paramount—yesterday's raw footage needs to become today’s viral social media clip. For live commerce and commercial production, the focus is granular: isolating specific product-showcase segments or cutting out specific hosts instantly. In legal and investigative fields, the priority is minimizing review times without missing critical contextual details. At the end of the day, every industry is trying to solve the same foundational problem: finding a 10-second needle in a multi-thousand-hour haystack.

When asked to summarize Heimdex's 2026 vision in a single sentence, Lee responded confidently: "Wherever video goes, Heimdex goes."

Kang outlined their immediate roadmap, which includes diversifying their international enterprise portfolio, launching a B2C platform, accelerating patent filings, and scaling their engineering talent.

"The upcoming B2C product will feature a different pricing model and more streamlined editing features tailored for creators," Kang explained. "We want a user to type a single natural language prompt and instantly retrieve the exact clip they're looking for. Our initial target audience isn't the casual consumer, but micro-influencers and YouTubers who manage substantial archives of personal footage."

Their approach to global expansion is equally calculated. The explosion of video data, compounding cloud storage costs, and tightening privacy regulations are universal corporate challenges. Lee confirmed that discussions with international production houses are already underway.

As video archives expand, organizations often find themselves drowned in data rather than empowered by it. Heimdex is cutting through that noise by eliminating the friction of data ingestion and deeply respecting the operational workflows of the industries they serve. Concluding the interview, Lee made it clear that Heimdex is thinking far beyond basic search.

"Video AI will continue to scale rapidly, and it will deeply permeate every facet of production," Lee said. "But whether you are building tools for AI editing or AI generation, every single one of those applications fundamentally requires video indexing first. We want to be the infrastructure—the backbone—that supports the entire future of the video AI industry."