Barcelona, Spain

Ivan Mikhnenkov

AI Engineer & Researcher

Portrait of Ivan Mikhnenkov

6+ years in computer vision and generative AI. I take models from research to production, owning both the AI features and the systems that keep them running.

Experience

  1. Ex-Human logo

    Ex-Human

    Feb 2022–present

    Lead Image Generation Engineer · since Oct 2023

    Previously Computer Vision Engineer / Researcher

    • Own the production AI system: serving multiple models across distributed GPU resources, APIs and business logic. Handle on-call incident response, monitoring and alerting.
    • Fine-tune image generation and editing models for apps with 500K+ monthly active users and 30M+ images a month. Build evaluation pipelines for image quality and identity consistency.
    • Optimize self-hosted diffusion and flow inference through quantization and compilation, reducing latency and memory use while preserving quality. Avoid tens of thousands of dollars per month in estimated API costs. Accelerated video inference to near real time (MiniMax H3).
    • Build LLM features and use vision-language models for moderation and quality estimation.
    • Ship product features independently, from prototype to deployment, and lead delivery across engineering and design. Guide iteration with Amplitude and PostHog.
  2. ZennoLab logo

    ZennoLab

    Dec 2021–Jul 2022

    Computer Vision Engineer / Researcher

    Built computer-vision services for browser automation and automated CAPTCHA solving.

  3. RT Labs logo

    RT Labs

    Feb 2020–Jul 2021

    Computer Vision Engineer

    Led satellite-based deforestation monitoring: designed custom neural architectures, trained and fine-tuned models, and built data and evaluation pipelines. Managed two engineers.

Independent research

TinyDiT

A 209M-parameter diffusion transformer trained from scratch on 4.2M images using flow matching. Data pipelines, training, evaluation and interactive inference tools.

Single-GPU training. Pretrained text encoder and autoencoder.

LLM post-trainingCode

PPO, GRPO & DPO

Implemented and compared three post-training methods in PyTorch. GRPO raised Qwen3.5-0.8B’s GSM8K test accuracy from 12.8% to 57.1%.

Full 1,319-question test set, greedy decoding. One run per method.

Independent products

I’ve independently built and tested AI product prototypes for video creation, multi-agent debates and e-commerce imagery, from concept to working features and market testing.

BeVisionary logo: a person looking at a star

BeVisionary2018 · First project

Designed, built and launched my first product: a goal-setting and journaling app.

Tech stack & tools

Tech stack
Python · PyTorch · FastAPI
Infrastructure
AWS · Docker · Redis
Monitoring & on-call
Prometheus · Grafana · Sentry · PagerDuty
Development tools
Claude Code · Codex · Cursor · Visual Studio Code
Product analytics
Amplitude · PostHog

Education

Higher School of Economics

Bachelor’s in Social Sciences
(Sociology / Quantitative Methods)

2016–2020

Art School · 2010–2014

English — full professional proficiency
Russian — native