Oscar Craven
Menu
Who’s looking?

The page reorders for you. Links keep your choice.

Oscar Craven, smiling, in glasses and a white shirt.

I build AI products people use, and the agent systems that ship and run them.

Hiring? My current role is below in five lines. The CV has the rest.

Download CV Send a message

Now: AI Implementation Consultant at AcmeAI & FX February 2026 to today

Show the five lines
What it is
Product, AI operations, hiring and partnerships for an AI film studio.
Who for
The studio and its artists, who make images and video with AI.
What I did
I was the first hire. I built Lupin, the studio’s AI image and video tool, and ran hiring with AI.
What happened
Artists now send work from Lupin straight into ShotGrid.
Next
Track which prompts lead to approved work.

The short version ends here.

Long read: what I did there

I was the first hire. The team is now 40+ people.

Headcount as of October 2026.

I built Lupin, the studio’s tool for AI images and video. What it does for artists:

  • ShippedSend to ShotGrid from LupinArtists pick the project, sequence and shot from live lists. Checks run before anything is sent. Refusals come back as a checklist they can read.
  • ShippedEvery output is filedEach output lands in the studio’s version store, with dailies rendered on the farm. File names pass pipeline checks unchanged.
  • ShippedPipeline media on the canvasArtists pull approved and working media, including plates, onto the canvas.
  • ShippedNo keys on artist machinesThe studio manages access to AI providers across three routes.

I ran hiring and recruitment with AI. Candidate search across portfolio sites, CV review and interview scheduling are automated.

I set up direct deals with model providers. They brought early access and free tokens for R&D. Going direct, not through image aggregators, cut model costs by 30 to 60%.

My estimate, 2026. It compares direct API prices for Seedance and Nano Banana Pro with aggregator prices.

Read the Lupin case study

Selected work

The bigger the block, the closer it is to the work I do now.

The funded build on a phone, with a language picker, an install button and accessibility settings.

LiveFunded build, launching December 2026

The Tenants’ Voice

Tenancy guidance in plain English, for renters in England.

Funding: £113,631 from the Social Housing Innovation Fund, April 2026, to a consortium led by NCRC.

New on staging, October 2026: you can talk to it out loud. See the voice experiment

The Lupin canvas: an image generator with its result, a list of 32 outputs and a naming node.

Shipped

Lupin

The studio’s tool for making images and video with AI. I built it.

Picture: Lupin desktop.

Google logo

AI compliance pilot at Google

A Vertex AI pilot that re-approved wrongly flagged items, for a Google advertising product.

78% re-approved automatically. Under 1% unsafe approvals.

Pilot results, 2024.

Can AI build a game?

An open benchmark: 580 builds by 9 models across 10 games, scored by 3 judges.

Published April 2026. Each build was scored by three AI judges from different model families.

Illustration: ten small retro game screens, such as a paddle game, a snake, falling blocks and a maze.

Before and after with behavioural science

An AI skill that rewrites a page and names the study behind each change.

With the skill, 30 of 30 checks passed. Without it, 13 of 30. Three mock briefs, graded by an AI model on 11 October 2026.

Ledgerly, a mock pricing page, after the rewrite: a new headline and three plans, with Pro in the middle marked “Recommended for agencies with 4+ clients”.

Companies I have built for

Brands and retailers whose pages carried my work, at Flixmedia and at Google.

Logos: Costco, Google, Sony, Dyson, AMD, Sage, Amazon, Bose, Samsung, P&G, dentsu, LEGO, John Lewis, Currys, Sainsbury’s, Braun, MediaMarkt, Mercado Libre, Argos and Darty.

What colleagues say

From people I worked with at Flixmedia and Google. Read all nine.

“Oscar is the kind of person you go to when you have a complex problem you cannot solve, and he will analyse it from A to Z and will have multiple solutions ready... He is the most technical PM I’ve had the pleasure of working with.”

Jane KuznecovaProject and Product Manager, Flixmedia

“His contributions evolved from execution to product innovation when he developed a novel solution concept for a strategic customer six months in.”

Ed Sala XifroProduct Manager, Google

“He’s a fast learner who not only quickly grasps new concepts but also dives into the heavy details to build a complete understanding. He then uses this deep knowledge to pioneer brand new approaches to old problems.”

André LourençoSolutions Architect, Google

Notes you can use

Short lessons from real projects. Each one ends with something to try. See all five notes.

The Tenants’ Voice

Why I said no to 13 things, in public

  • Tenants made 28 asks. They became 9 features.
  • 13 things were declined, each with a public reason.
  • No live chat, said the scope. I built a voice version anyway on staging. It is the one no I took back.

Try this: write the reason down. People accept a no they can read.

Workshop: summer 2026. Scope of works: 2026. Voice experiment: staging, October 2026.

See all 13 reasons

The Tenants’ Voice

The stack mattered as much as the model

  • Same model, same 50 test cases, two set-ups.
  • 2.5 Flash scored 3.64 on staging and 2.66 on production.
  • So about one point came from retrieval and the prompt.

Try this: fix your retrieval and prompt before you pay for a bigger model.

Scores out of 5. Blind review by five Claude reviewers, 2 October 2026.

See how we tested

Lupin

From a canvas straight into ShotGrid: what the pipeline checks first

  • Artists pick project, sequence and shot from live lists.
  • Checks run before anything leaves the canvas.
  • A refusal comes back as a checklist, not an error code.

Try this: check early, and explain every refusal in plain words.

Read the note

How I work

I run a small team of AI coding agents. I still read every change before it merges.

Build
Each agent takes one task on its own branch. Nothing merges until I have read it.
Learn
I send links from my phone. Agents file them and send back a short brief.
Improve
Each session ends with what I learned and what is still open.
Illustration: a desk seen from above, with a laptop, a phone, a notebook with a checklist and a coffee.
Where the three loops happen: laptop, phone and notebook.
Film: how one request becomes a reviewed change. Captions are on.
See how I work

Lab

Small tools I built for my own day. Try them, or read the code. See the whole Lab.

Open Loops: one list of who needs you and who you are waiting on.
One list, every morning.

Open Loops

Lists what you asked people for on Slack and Gmail, and what they are waiting on you for. Then it drafts the follow-up in your own words.

It uses the Claude, ChatGPT or Grok plan you already have.

See the code
Plainspeak explaining a message in plain English, in a small panel beside the pointer.
Plainspeak, on fictional messages.

Plainspeak

A Mac menu-bar app. It explains or corrects the message under your pointer, in plain English. It never sends anything.

It runs Claude Haiku on your own Claude plan.

See the code
Slipway: one branch’s files on the left, its agent terminals in the middle, and a console and dev server below.
Slipway in use. Terminal text is blurred.

Slipway

A desktop app that gives each Git branch its own worktree, terminals and GitHub status. It shows what the coding agents in each one are doing.

I built it for a faster working day.

Install Slipway

Send me a message

Tell me what you are working on. I read every message.

What are you working on, and how could I help?

An email address. Leave it blank if you do not need a reply.