CodeSpeak kills all boilerplate so that humans only interact with meaningful code. If some code can be reliably generated by a machine, we can replace it with a concise description in plain English. We all know that most code in this world is not very dense in meaning, so a CodeSpeak codebase will be a lot shorter and easier to read and modify than usual. Think “CRUD REST API for Contacts and Addresses” instead of 50 lines of trivial code.

This approach distils codebases down to their essential meaning: all the fluff is gone and every line is worth reading. There’s much less code, and it’s much easier to understand, navigate, and evolve. While we also use LLMs and agents to generate code, CodeSpeak is not a vibe coding tool. It’s more like a higher-level programming language that understands English and uses an LLM as a library.

CodeSpeak vs Vibe coding. Coding agents understand human language but it gets discarded as soon as your session is over. Your teammates are left to read machine-generated code, without access to how you described it in plain English. We end up talking to machines in human language, while communicating with other humans in machine language (code). In CodeSpeak, the human-language intent is never lost, and everyone communicates on the same level of abstraction.

What you will be working on

• Code generation pipeline: turn English descriptions/specs into code,
• Boilerplate elimination pipeline: turn trivial code into English descriptions,
• Evaluations for both pipelines,
• Code generation stability and correctness: make sure the generated code is equivalent to the boilerplate it is replacing,
• Real-time feedback (as you type): when users edit English descriptions, we must be helping them on-the-fly, like a smart IDE would,
• Code merge pipeline: when users merge git branches, they can merge English, but we have to merge generated code for them,
• Latency improvements: we want to minimise waiting times to keep the users in the flow state.

Ideal candidates

Smart, curious, self-driven applied researchers who believe in our mission.

Green flags include hands-on experience with:

• Solid foundation from classic ML to modern neural networks,
• Expertise in LLMs: serving, fine-tuning (SFT, LoRA), quality evaluation, and designing custom benchmarks,
• Experience designing and evaluating agentic/multi-agent systems from 0 to 1,
• Strong skills in inference optimization and agentic workflow tuning: quantization, batching, KV-cache, context engineering, and overall reducing latency and cost.

Current team

You’ll be working alongside excellent engineers and AI researchers from Kotlin/JetBrains, DataBricks, Isomorphic Labs, and Google. The team is in the startup mode, so we encourage ownership, exploration, and cross-discipline work. We are focused on delivering results that make users happy.

---