Vetle
Gangeskar

Applied AI & Creative Technology.

Model · System · Production · Performance

Oslo, Norway
Building across the Nordics

I build the systems between new AI models and real creative performance. Test the model, find where it breaks, build the workflow around what survives, put it into production, measure what it did, feed that back in.

See the systems Get in touch

  1. 20,000+

    Subscription sales

    Across creative work and systems I’ve worked on.

  2. Hundreds

    Creatives tested monthly

    Short-form video, UGC, static and AI-generated.

  3. 4

    Languages in production

    Norwegian, Swedish, Danish, Finnish — from one creative system.

  4. Private beta

    LURK

    Software I built. Preparing for launch.

01 Systems

I built the infrastructure I wanted to use.

Competitive intelligence sits in one tool. Generative models in another. Production somewhere else. Ad performance somewhere else again. Nothing talks to anything, so the loop never closes. LURK is my attempt to close it.

LURK

Creative intelligence infrastructure, built from scratch

  • Python
  • FastAPI
  • PostgreSQL
  • Redis
  • Celery

Today LURK collects advertising creative at scale, structures it, ranks it and surfaces the patterns underneath — a running service with an ingestion pipeline, a database, scheduled workers and ranking logic.

Ingestion runs through a priority chain — the official Graph API first, lighter methods as fallback — so collection degrades gracefully instead of breaking when one route closes. Scheduled Celery workers re-scan a watchlist and flag when a competitor’s creative changes. That is the part that makes it infrastructure rather than a research tool: the corpus stays current whether or not anyone opens it.

The ranking is deliberately honest about its own limits. It reflects delivery signals and ad age — not revenue, not ROAS — and the interface says so on the page. A score that overclaims is worse than no score.

Role
Sole builder — product, backend, ranking logic
Runs
Scheduled ingestion & analysis workers
Status
Private beta — preparing for launch
Started
2025

How it gets built

I define the architecture and the tradeoffs, build with modern AI development tools, then debug, deploy and iterate until it holds under real load. The ingestion chain has a fallback because I hit that failure in production — not because a tutorial suggested one. What matters is the distance from idea to a running system, and how fast the loop closes after it breaks.

Where it is going — not built yet

Today LURK reads the market. The system it is becoming closes the rest of the loop: from a recurring creative pattern, to which generative model and workflow can actually reproduce it, to what that costs per usable output, to how the creative performed once deployed — and back into the next selection. Ranking and ingestion exist. The model-selection and performance-feedback layers are the work ahead.

  1. Creative intelligence
  2. Model & workflow fit
  3. Generation
  4. Deployment
  5. Performance
  6. Learning

Solid = running today. Outlined = being built.

Architecture

  1. Discover
  2. Ingest
  3. Structure
  4. Analyze
  5. Rank
  6. Experiment

The last step is the only one that is not automated. Everything before it exists to make that decision better than a guess.

Try LURK

↺ The method

One loop. Three surfaces.

The projects below look unrelated. They are the same method pointed at different problems: build the thing, put it under real load, measure honestly, change one variable, repeat.

  1. 01

    Build

    Write the system. Software, workflow or edit — whatever the problem actually needs.

  2. 02

    Test

    Put it under production load, where being wrong has a cost.

  3. 03

    Measure

    Name the failure mode. “It felt off” is not a measurement.

  4. 04

    Learn

    Change one variable, so the next result means something.

  • Software

    LURK — a creative intelligence platform I built and run.

  • Models

    Model Lab — frontier video and image models, judged on production output.

  • Distribution

    Production — Norwegian Lab and Avalanche Safety App, where it has to hold.

02 Model Lab

Frontier models, judged on what survives production.

Generating one good clip is not the interesting problem. The interesting problem is knowing what breaks, why it breaks, which model to reach for, and how to build a workflow that produces usable output on a deadline.

New image and video models land constantly, and each one changes what is worth attempting. So I test them against each other on real briefs rather than demo prompts — same script, same constraints, same delivery date. That turns a subjective impression into something closer to an evaluation.

What I care about is the usable-output rate: of the generations a model produces for a real brief, how many survive an edit. A model that makes one spectacular shot and nine unusable ones loses to a model that makes six serviceable ones.

What I evaluate

  • 01Identity consistency across shots
  • 02Temporal coherence — drift, morphing, flicker
  • 03Prompt adherence and controllability
  • 04Motion quality, and whether camera movement is motivated
  • 05Product consistency — packaging, logos, label text
  • 06Lip sync and speech alignment
  • 07Artifact rate — hands, edges, text, physics
  • 08Editing headroom — does it cut, or only look good alone
  • 09Generation speed and iteration cost
  • 10Usable-output rate against a real brief

How a run is recorded

Every model that reaches a real brief gets logged the same way: sample size, how many generations survived an edit, how many were discarded, cost per usable output, generation time, and the failure mode worth remembering. That is what makes two models comparable. A single spectacular frame does not.

  • Model
  • Task
  • Samples
  • Usable %
  • Failure %
  • Cost per usable
  • Generation time
  • Best for
  • Failure mode

How a test runs

Step three is the one that matters. What a model does when it breaks tells you more about whether you can build on it than any hero shot.

01

Reference

One identity. One product. Constraints that cannot move.

02

First pass

Convincing until temporal consistency is inspected.

03

Failure

Identity drifts through the camera turn; the product changes between frames.

04

Iteration

Changed reference strategy, shot design and generation workflow.

05

Shipped

Stable enough for production, and repeatable on the next brief.

Currently working in

  • Higgsfield
  • Seedance
  • Premiere Pro
  • DaVinci Resolve
  • Prompt & context engineering

03 Production

Where the models meet a deadline.

A workflow that works in a demo and a workflow that survives a release calendar are different objects. This is where I find out which one I have built.

Norwegian Lab

Performance marketing & AI creative strategy · 2024—present

Norwegian Lab is my real-world AI laboratory. It is where new generative models meet the constraints that actually decide whether they are useful: speed, consistency, localization, creative fatigue and iteration.

I take it the whole way — insight, hook and script, through production and editing, into analysis and the next round. Hundreds of creatives tested every month, in four languages in production — Norwegian, Swedish, Danish and Finnish. At that volume taste alone stops scaling: you need a named reason a sequence worked, or you relearn the same lesson every week.

A large and growing share of that output is AI-generated or AI-assisted. I build the production workflows that raise output, localization capacity and testing speed while holding product and brand consistency — and every AI-generated placement ships with an on-screen disclosure.

20,000+ subscription sales across creative work and systems I’ve worked on.

Scale
Hundreds of creatives tested every month
Languages
4 in production — Norwegian, Swedish, Danish, Finnish
Formats
Short-form video · UGC · AI-generated · static
Does
Insight → script → generate → edit → test → iterate
Norwegian Lab

Avalanche Safety App

Growth & distribution

A Norwegian safety app that helps locate a buried person using GPS, Bluetooth and sound — a supplement to a transceiver, never a replacement. I work across launch, creative, growth and distribution, taking the product from early adoption into scale — currently around 1,000 paying users.

Safety products have an unusual distribution problem: the best product does not win automatically. It wins when it is the one people already trust — seen in the lift, in the ski bus, recommended by a guide. That makes it a credibility problem before it is a media-buying problem, and it is the clearest case I work on where a technically good thing still needs a system around it to reach anyone.

Traction
~1,000 paying users
Product
Avalanche location aid — GPS, Bluetooth, audio
Market
Scandinavia, expanding into Europe
My focus
Launch, creative, growth and distribution
Why me
I ski this terrain. I know who the audience trusts.
App Store About the app

04 CV

The short version.

  1. 2024 — present

    Performance Marketing & AI Creative Strategy

    Norwegian Lab · Norway

    • Take a concept from insight and script through production and editing, into analysis and the next round.
    • Test hundreds of creatives every month across short-form video, UGC, static and AI-generated formats, in four languages in production — Norwegian, Swedish, Danish and Finnish.
    • Continuously evaluate new video and image generation models against real briefs, and decide which go into production.
    • Build AI-assisted production workflows that raise output, localization capacity and testing speed while holding brand consistency.
    • Use LURK, my own platform, daily to set the brief before production starts.
  2. 2025 — present

    Founder — LURK / Creative Intelligence Engine

    Oslo

    • Designed, built and operate a platform that collects, structures, analyzes and ranks advertising creative.
    • Python, FastAPI, PostgreSQL, Redis and Celery — ingestion pipeline, scheduled workers, ranking logic and product.
    • Sole builder across backend, data pipelines, recommendation logic and product strategy.
  3. Ongoing

    Growth & Distribution

    Avalanche Safety App · Norway

    • Launch, creative, growth and distribution for a consumer safety product at ~1,000 paying users, ahead of and through the avalanche season.
    • Creative production, channel testing and measurement across paid and organic.
  4. Earlier

    Background

    • Sateba Norway, Hoenefoss — general worker on the production floor of a precast concrete plant, 2021–2023. Shift work alongside school and competition.
    • Bachelor in Art Direction, specialization in Marketing Communication — Kristiania / Westerdals, Oslo.
    • Ringerike videregående skole — upper secondary, sciences and mathematics.
    • The Norwegian Army — military service, 2023.

↘ Before software

I learned iteration in the mountains first.

I compete in freeride skiing at international level. It is the same loop as the work, with a shorter feedback delay and less forgiving consequences: read the terrain, commit to a line, watch what actually happened, adjust.

  • 2024

    Qualified for the FWT Challenger Series

    The only Norwegian rider in the Freeride World Tour Challenger Series — the tier leading to the Freeride World Tour and to the first FIS World Championship in freeride skiing.

  • 2023

    Winner — Norgescup, Hemsedal

    First senior Norwegian Cup win, and the youngest in the field.

  • 2022

    Winner — Freeride World Tour Junior

    The junior world tour title in freeride skiing.

  • 2018

    First descent — Mørkgonga

    Among the first to ski the landmark line, as a junior.

Freeride gives you one run. You inspect the face from the bottom, build the line in your head, and then it holds up or it does not — in front of judges, once. That is a useful thing to have practised before a career where the honest question is always the same: did that actually work, or do I just like it?

I grew up in Røldal, a mountain village in western Norway of a few hundred people, where the nearest anything is an hour of road away. Places like that teach one reflex: if you want a thing to exist, you are probably the one who has to make it. I built my own skis out of wood before I built my own software.

The ski career A season on film

Building something difficult with AI?

I am always interested in talking to people working on generative media, AI systems, and the messy part where models meet the real world. Whether or not there is a role attached.

vetle@gangeskar.com LinkedIn