London, United Kingdom

Open to Staff / Principal / Lead engineering roles — AI & ML platforms, data infrastructure and distributed backends. London or remote. yauhen.bichel@gmail.com · LinkedIn · Hugging Face · Medium

Software engineer with 14+ years building agentic AI applications, ML systems, big data pipelines, distributed backends and cloud infrastructure. I build systems from scratch — from CNN training and TensorFlow serving, through Kafka event pipelines and Spark/HDFS batch computation, to the Terraform that provisions all of it.

Java · Go · Python · Node.js · C#/.NET on AWS · GCP · Azure

Currently a core open-source contributor to OpenSRE (Apache-2.0 framework for AI SRE agents, 10k+ ★) and co-founder of MoleCare, a skin-lesion monitoring app published on the NHS App Library.

molecare-mcp — MCP server giving Claude and other clients educational dermatology knowledge: lesion terminology, ABCDE criteria, SNOMED CT to ICD-10 mapping. Runs with no credentials.

npx -y molecare-mcp

molecare-ml — the melanoma classifier behind MoleCare, published with a model card that states plainly what was not measured. Its write-up on why 94% accuracy means very little explains the gap, and the open bias evaluation issue is the problem I would most like help with: performance across Fitzpatrick skin types is unmeasured, and dermoscopic datasets under-represent darker skin.

molecare-desktop · molecare-skin-llm — the Electron desktop build, and a LoRA-tuned educational Q&A model behind a deterministic safety harness.

py-harness — four jobs against a local model on your own machine: ask a question, write a test, fix a bug, add one small function. It only touches the folder you point it at.

pip install py-harness-cli

moe-fit — answers "will this mixture-of-experts model run on my machine, and how fast?" from the model's index, before downloading it. DeepSeek V3.1 is 405 GB, but only 11.4 GB of that is needed by every token, so it does run on a 128 GB box — at about a tenth of a token a second, which is an hour for a paragraph. You learn that in three seconds rather than after a 405 GB download.

pip install moe-fit
moefit plan <model.gguf or a URL to one>

merge-cheer — a GitHub Action that comments a G-rated GIF when a pull request merges. No Giphy key, no secrets, and it ships its own loops so it works on a new repository with the default token.

readme-contributors — a GitHub Action that draws the contributors wall further down this page. Circular avatars, no table, bots omitted.

None of it is a medical device. None of it diagnoses anything.

Local AI and robotics

• strix-halo-jax — JAX and MuJoCo MJX on an AMD Ryzen AI MAX iGPU.
• belarusian-tts — Self-hostable Belarusian text-to-speech.
• humanoid-companion — A small humanoid robot that walks with a reinforcement-learning policy.

Infrastructure and tooling

• terraform-aws-parameter-store — Terraform module for AWS Parameter Store.
• terraform-aws-eventbridge-batch-trigger — Terraform module for EventBridge-triggered AWS Batch jobs.
• terraform-aws-batch-job-revision — Terraform module for AWS Batch job revisions.
• github-action-terraform-output — GitHub Action that reads Terraform output.

Services and applications

• demo-openlr-mapbox-data — Decoding OpenLR location references against Mapbox data.
• Services-with-Mesh-and-gRPC — Microservices over gRPC behind a service mesh.
• Text-to-image-RESTful-service — REST API for Stable Diffusion image generation.
• Retry-with-resilience4j — Retry policies with resilience4j.
• AudioPlayer — React Native audio player for iOS.

Over 1,000 merged pull requests across roughly 97 repositories, mostly on the projects above and in the MoleCare organisation. I also contribute to OpenSRE, an Apache-2.0 open-source project for AI SRE agents.

Thank you to everyone who has helped.

Filled from GitHub commits (bots omitted). Live demo: readme-contributors.