Miju Labs

← All buyers · Recraft

P

Pavel Ostyakov

Head of AI, Recraft · Recraft

medium confidenceTier A · Talk to firstImage, video & audio modelsimage

Runs Recraft's AI team — around eleven people per The Org — and is the named technical voice on how Recraft V4 was trained. He is the most actionable working-level buyer in this batch: a small, London-based model team that owns training, evaluation and, by implication, the annotation programme Recraft advertises for in the same Engineering org.

Background

Kaggle Grandmaster turned applied-research lead — his LinkedIn headline carries the Grandmaster credential, which says something about how he thinks: metrics, leaderboards, measurable wins. He leads model training at a company that builds ~20B-parameter image models from scratch rather than fine-tuning, on Nebius infrastructure and NVIDIA HGX B200 clusters.

His public statements are about scaling and precision — bigger models give better image quality; model size means more than parameter count; simple code lets you switch GPUs without special work. Notably absent from his public record is any account of how design quality is measured, which is the gap a taste-data vendor exists to fill.

What they run now

Training Recraft's foundation models: V4 as a ground-up rebuild, bfloat16 for the main model, FP8 for the text encoder, NVFP4 experiments for inference, validated on B200 early access. The stated quality target is visual judgement and brand-level quality for professional designers and marketers — a human-judgement target with no published human-judgement methodology. Recraft's annotation operation (MTurk, Toloka, Clickworker) sits in the same Engineering organisation in London.

What they buy

Strong inference plus direct job evidence in his org: crowd annotation via MTurk, Toloka and Clickworker (captioning, tagging, QC) managed by a Data Annotation Manager role in Engineering, London; and — unevidenced but implied by the V4 positioning — designer preference data and aesthetic evaluation. The crowd-platform mix is the tell: it covers volume labelling but cannot cover professional design judgement.

Career

  1. presentHead of AI, Recraft — ~11 direct reports per The Org
  2. —Grandmaster (competitions), Kaggle — Credential carried in his LinkedIn headline

On the record

scale and image quality · 2026-02

Says that for image generation, the bigger the model the better the quality. source

what model size means · 2026-02

Argues model size means more than just parameter count. source

portability · 2026-02

Notes that simple code lets you move to new GPUs without doing anything special. source

Interviews & talks

Why it matters here

The best working-level target in this batch. He owns training and evaluation at a company whose entire pitch is design taste, his team already runs annotation as an operational function, and the supply chain he uses (generic crowd platforms) is structurally unable to deliver the judgement his marketing claims. He is in London, the team is small, and he responds to measurable arguments.

How to reach

Speculation: send a measurement artefact, not a deck. Run a small blind panel of professional designers comparing Recraft V4 against two rivals on layout, typography and client-usability, report panel size and inter-rater agreement, and send it with the offer to scale it. Frame it as the missing number next to his scaling numbers. A secondary hook: Ideogram published exactly this kind of evidence for its 4.0 release and Recraft did not — a competitive argument that lands with a Kaggle Grandmaster.

What we could not establish
  • Whether Recraft uses any specialist (non-crowd) vendor for design judgement today.
  • Whether the London Data Annotation Manager role reports to him — the posting is in Engineering but gives no reporting line.
  • No X handle verified.
  • No talks, papers or podcast appearances found beyond the Nebius case study.

Why they are on the map

Named Head of AI in Nebius Recraft V4 case study; runs model training. source — evidence dated 2026-02

Needs the data for a model; writes the spec.

How Recraft buys data

Recraft V4 (2026-02-17) was 'developed in close collaboration with designers, tuning it around design aesthetics and professional expectations'; no methodology or vendor disclosed. Trains own foundation models on Nebius (V4 case study, early 2026). 350M+ images generated on platform provide in-house preference signal.

CEO Anna Veronika Dorogush (ex-CatBoost lead) with Head of AI Pavel Ostyakov running ~11 reports; small Head of Design function under Ivan Olenkevich.

Sources

  1. Nebius customer story: Recraft V4 (names him Head of AI, quotes him)
  2. LinkedIn: Pavel Ostyakov — Head of AI @ Recraft, Kaggle GrandMaster
  3. The Org: Recraft org chart (~11 reports)
  4. Recraft Data Annotation Manager job, London, Engineering