Kerim Yalchin

Software engineer based in London.

Previously at Whitespace / Verisk and Acue. Now working on developer tools, desktop applications and applied ML.

Professional experience

2022—Now

Independent

Software Engineer

Building and maintaining developer tooling, distributed desktop systems, simulation software and GPU-accelerated ML systems. Work includes Tethoq, NH Trainer, Shape Vectorizer, Klipt and glTF-to-JSON.

2022

Whitespace / Verisk

Software Engineer · Contract

Improved page loading by 10× through a memory-management change in a core algorithm, then shipped TypeScript and Angular features and resolved production issues for client-facing software.

2021—22

Acue

Software Engineer · Contract

Designed the SQL model and AWS environment for a Java/Spring sports platform, including EC2 and Cognito, while leading the resolution of delivery-blocking stories in a CI/CD team.

Selected projects

Developer tools

Tethoq

A provider-neutral workspace for local and remote coding-agent sessions.

Built across Electron, React and Flutter, with a TypeScript bridge for provider integrations, remote sessions and approvals. It brings coding agents, browser tools and multi-agent work into one interface.

TypeScript
React
Electron
Flutter
WebSockets

tethoq.com View source Watch demo · 31 sec

Tethoq desktop dashboard showing synthetic coding-agent tasks

Tethoq mobile dashboard showing sample coding tasks, approvals and a connected desktop

Neural networks & simulation

NH Trainer

A browser combat simulator powered by custom-designed neural-network opponents.

Designed and trained the opponents alongside the combat simulator, from the training pipeline and policy evaluation through to browser inference and the playable interface.

TypeScript
React
Three.js
Java
Reinforcement learning

Try NH Trainer View source

NH Trainer browser client during a combat practice session

GPU computing & graphics

Shape Vectorizer

A GPU raster-to-vector optimizer that redraws images as editable SVG.

Fits hundreds of layered vector shapes to a source image using a custom CUDA renderer and analytic gradients, with a portable PyTorch fallback. The result remains editable as SVG.

Python
PyTorch
CUDA
CuPy
SVG

View source

Original image, vector reconstruction and error comparison from Shape Vectorizer

Desktop software

Klipt

A focused desktop tool for turning long videos into precise clips.

An Electron application that coordinates yt-dlp and FFmpeg behind a small, usable interface. It supports source quality selection, exact time ranges and self-healing dependency updates across YouTube, Twitch and Kick.

Electron
Node.js
FFmpeg
yt-dlp
HTML/CSS

View source

Klipt desktop application showing clip timing and quality controls

3D developer tools

glTF to JSON

A one-file converter for getting animated 3D assets into constrained game runtimes.

A Python GUI and CLI that converts glTF/GLB assets into compact, integer-quantized JSON while preserving vertex order, hierarchical skinning and stitched animation clips.