KLADON Book a call
K-INFO / 01·STUDIO·FOUNDED 2026

A studio
run like a research lab.

/ 01 Who we are WHY THIS STUDIO EXISTS

Kladon was founded by people who spent more than a decade inside OTT streaming. Running product, watching the catalog, fighting CDNs, shipping the apps.

Generative AI went from research curiosity to production reality between 2023 and 2026. Media companies have struggled to absorb it. Buying point tools, building one-off integrations, hiring ML engineers who don’t know what SMPTE timecode is.

Kladon is the studio we wished existed when we were on the buyer side. A small team that understands the catalog, the timecode, the rights chain, and the difference between an interesting demo and a feature your ad sales team can sell against.

We’re AI-native. We use AI in the work, and we ship AI as the work. We don’t train foundation models from scratch. We compose the existing ones into pipelines that solve a specific editorial, operational, or monetization problem at catalog scale.

Deliverables aren’t slide decks. They’re deploy URLs, TestFlight builds, Play Store listings, operational dashboards, and the line on the P&L that pays for the engagement twice over.

/ 02 Team 05 OPERATORS

A small team, by design.

Five people. Engagements are run by senior operators, not by interns with a pipeline diagram. The team grows only when an engagement asks for it. Most of us go by handles in chat; full names show up on the contracts.

/ STUDIO LEAD

@wjkoo

15 years inside OTT streaming product. Has run the catalog side, the apps side, and the commercial side at Asian-content operators. Sells the work, sits in the kickoff, stays in the weekly. Knows the buyer because they were the buyer.

  • Studio direction
  • Product
  • Commercial
  • Domain
/ PRACTICE · ML

@mookai

Vision lead. Segmentation, tracking, scene understanding, diffusion-based inpainting. Spends most days inside SAM3, YOLO26, and a stack of bespoke fine-tunes. The person who decides whether your edge case needs a new model or a better prompt.

  • SAM3 · YOLO26
  • Diffusion · roto
  • Eval harness
/ PRACTICE · ML

@shenru

Speech and multimodal lead. ASR for Korean, Japanese, Mandarin, and English. LLM orchestration for scene captions, brand-safety, and recommendation embeddings. Owns the dubbing and lipsync stack end to end.

  • Whisper · NeMo
  • CLIP · SigLIP
  • Dub & lipsync
/ PRACTICE · INFRA

@nguyenj

Backend, GPU orchestration, throughput. Has shipped under SLA and has opinions about timeouts. Owns the eval harness, the queueing layer, and the cost-per-source-hour number that everything else has to live inside.

  • FastAPI · Temporal
  • Ray · vLLM
  • GPU orchestration
/ PRACTICE · VIDEO ENG

@t-nakamura

Video pipeline and mobile engineering. FFmpeg, codecs, CMAF, IMSC1 in one direction; CoreML, MLX, NNAPI, and on-device runtimes in the other. The reason on-device clips look the same as cloud-rendered ones.

  • FFmpeg · CMAF
  • Swift · Kotlin
  • CoreML · NNAPI
/ 03 Process FROM SIGNAL TO SHIP

A loop, not a waterfall.

Diagnose, Design, Deploy, Operate. The same loop runs every engagement, scaled between two weeks and twelve months. We compress the handoff between research and production by keeping the team small enough that there isn’t one.

01
Diagnose
Two-week scoping sprint. Ingest a sample catalog, audit the current stack, interview the operators and the ad sales team. Output: a written brief and a target P&L line.
$15–25K · WK 1–2
02
Design
Architecture, data flow, model selection, evaluation criteria. The eval harness is written before the inference pipeline, not after.
WK 2–4
03
Deploy
First production output by week 4. Iteration is measured against editor sign-off rate, CPM lift, fill, watch-time, or whatever the target line was at week 1.
WK 4–10
04
Operate
Monthly retainer or a clean hand-off package. The field moves week to week, so we keep models, prompts, and weights current.
WK 10+ · MONTHLY
/ 04 Principles 05 OPERATING RULES
/ 01
Models are components.
No single model wins everything. We pick the right tool for the task (open weights, hosted API, or a fine-tune) and swap it without drama when something better lands.
/ 02
Editorial sign-off, not autopilot.
Brand-grade media still needs a human in the loop. We build editor-assist tools where automation does the bulk and the editor stays in command.
/ 03
Deploy on day one.
Every engagement starts with a deploy URL or a TestFlight build. Progress shows up in build numbers, not Jira tickets.
/ 04
Rights, watermarks, provenance.
Media AI without C2PA, music licensing, and rights metadata is a lawsuit waiting to happen. We build compliance in from day one.
/ 05
Money has to follow the work.
Every engagement maps to a P&L line: ad CPMs, fill rate, ARPU, watch-time, retention, marketing throughput, compliance cost avoided. If we can’t name the line, we don’t take the work.
/ 05 FAQ WHAT WE GET ASKED
/ 01
Are you an agency, a studio, or a software company?
All three, deliberately. We sell engagements like an agency, we ship products you can buy like a studio, and we run our own software in production (Scene Intelligence Engine and Kladon Clips). The mix keeps the work honest.
/ 02
Do you train your own foundation models?
No. We use the best open and hosted models available (SAM3, YOLO26, Whisper, Claude, Gemini, MLX) and fine-tune where it materially moves quality. Training a foundation model from scratch is the wrong way to spend a media-studio budget.
/ 03
How does the work pay for itself?
Most engagements target a specific P&L line: CPM lift on contextual inventory, fill on under-sold AVOD pods, social-driven trial conversion, watch-time and SVOD retention, marketing throughput per editor, or compliance cost avoided on EU AI Act and accessibility deliveries. We name the line in week 1 and report against it weekly.
/ 04
Do you take equity instead of cash?
Occasionally, for early-stage media-tech companies where we’re closely aligned. Most engagements are paid in cash, USD.
/ 05
What languages and locales do you cover?
English, Korean, Japanese, Mandarin, and Spanish out of the box. Vietnamese, Indonesian, and Thai on request. Workflows are tuned for Korean, Japanese, and Mandarin catalogs, including subtitles in IMSC1 and SMPTE-TT.
/ 06
How do you handle compliance?
Every output is C2PA-signed, optionally watermarked, with rights metadata at frame granularity. Music licensing runs through the ASCAP, BMI, JASRAC, KOMCA, and MCSC flows. We’re tracking EU AI Act Article 50 (August 2026) closely.
/ 07
Where are you based?
HQ in Boston, Massachusetts. A West Coast presence in Los Angeles, mostly for the studio and streaming clients there, and a Shanghai office that handles APAC partners and on-the-ground content work. Clients are mostly in the U.S. and across East and Southeast Asia. Engagements run hybrid; we travel for kickoff and pre-launch.

Bring us a hard problem.
We bring the frames.

Start a project See the studio