Government technology has failed citizens, veterans, public servants, and service members for decades.

Kaizen powers America’s application layer. We replace the legacy systems federal agencies run on with modern software, built with craft. The ticketing platform a resident uses to visit a presidential library. The marketplace DoW uses to procure counter-drone systems to protect lives and critical infrastructure. The system a Veteran uses to access benefits. We build systems of record, of national consequence. Our AI-native modules and harness allow us to do so in weeks, not years.

This year, our revenue has grown 35x. Next year, our software will serve over 100 million Americans across dozens of federal agencies.

Founded in 2022 and based in New York City and DC, Kaizen has raised $35 million from NEA, a16z, Accel, 776, and Carpenter Capital.

Location

New York, NY (Hybrid). In office Monday to Thursday. Candidates must live in or be able to commute to New York City.

The Role

Most companies building with AI connect someone else's models to their software. Kaizen's AI-native platform turns requirements into working applications, and today much of that intelligence comes from general-purpose models that carry capabilities we don't need, lack knowledge we do, and leave us and our customers paying frontier-model token costs for intelligence we don't control. This role changes that. You'll fine-tune, adapt and distill open-weight models into specialized intelligence Kaizen owns, and prove where it beats the general-purpose alternative. The goal isn't the biggest model or the most papers: it's public-sector software that works better, costs less to run, and is harder to replicate.

What You'll Do

Fine-tune, adapt and distill open-weight models for requirements interpretation, code generation, repository understanding, structured output and software validation.

Own the experiment loop: curate datasets, build training pipelines, and test architectures and techniques until a promising idea becomes a working model.

Build evaluations on real software engineering tasks that set baselines, expose failure modes, and separate real gains from benchmark noise.

Work with platform engineers to find where current models fall short, then build the capabilities that close those gaps and ship them into production.

Experiment with supervised fine-tuning, synthetic data, preference optimization and reinforcement learning to improve model behavior on specialized tasks.

Deliver applications built around each customer's needs without frontier-model token spend, so public budgets go to residents, not inference.

Lay the groundwork (datasets, evals, research practices) for Kaizen to eventually train its own models from scratch.

What You'll Bring

Hands-on experience training, fine-tuning or adapting language models. You've worked with model weights, not just model APIs or prompts.

Strong software engineering fundamentals. You build reliable training pipelines, debug complex systems, and write code other engineers build on.

Strong Python and PyTorch, with a working understanding of transformer architectures, tokenization and language model training.

Something you've built that shows real model engineering ability: a fine-tuned model, an original training approach, or a meaningful open-source ML contribution. Industry, research or independent work all count; demonstrated ability matters more than years or credentials.

Strong Candidates May Also...

Have fine-tuned models for code generation, program synthesis or repository understanding.

Design experiments and evaluation sets that tell a meaningful improvement from noise or overfitting.

Bring experience with synthetic data, distillation, quantization, or preference optimization and RL with verifiable rewards.

Have contributed to open-source model development, published applied ML research, or reproduced significant research results.

Don't Apply If...

Your model experience is prompt engineering and API calls.

You'd rather publish a result than ship a model people use every day.