Software Engineer: Self Improvement

Big Wombat LLC; New York (NY); Remote; USD 215,000 - 325,000; Full time

Benefits offered by this job

Equity

Job summary

Yaklabs is building a desktop AI application that turns knowledge workers into AI power users. We’re hiring for an engineering role that will own the self-improvement loop in Bonsai, with remote work in US or EU timezones and a full-time schedule in the US.

You’ll work across learning from use, right improvements, evaluation, and user-visible sharing. This founding-team role integrates product sense, strong writing, and hands-on building of retrieval, evals, and agent loops, with in-person NYC

Qualifications

• Experience building with LLMs including retrieval, evals, and agent loops.
• Excellent writing and ability to express instructions clearly.
• Strong product instincts and ability to make complex systems usable.
• Ability to design evaluations that reveal truth and avoid hype.

Responsibilities

• Build the end-to-end self-improvement loop from evidence to improvements.
• Develop the skill runtime: discovery, retrieval, loading, and composition.
• Create pipelines turning session history into skills and memories.
• Design user experience for non-technical users to review and steer learned stuff.
• Build evaluation systems with task suites and production signals.
• Own sharing, distribution, versioning, and provenance.

Skills

LLM workflows

Retrieval

Evals

Product sense

Strong writing

Job description

Engineering, Remote (US or EU timezones), Full-time

You turn Bonsai’s compounding self improvement loop into a product that non-technical people can use. As someone works, Bonsai learns their context, procedures, judgment, and taste, and uses that to help with the next task. You own how real work becomes improvements, how we prove those improvements help, and how people see and control them without writing prompts or configuring an agent.

Problems you’ll work on

These are examples, not lanes. You will move between them, while taking real ownership of the ones that match your interests.

• Learning from use: Every session is evidence of how someone likes their work done. How do you spot a repeated workflow or a correction, turn it into an improvement, and propose it at a moment the person will say yes, without burying them in suggestions they didn’t ask for?
• The right improvement at the right time: One person may end up with hundreds of skills and memories. How does an agent pick the few that apply to this task, combine them, and ignore the rest, inside a finite context window?
• Proving an improvement helped: A skill or a memory that reads well can still make outcomes worse. How do you evaluate improvements against real tasks, notice when one goes stale or starts hurting, and fix or retire it?
• Improvement people can see and steer: Most people will never read a prompt or edit a config file. How do you show what Bonsai learned, let someone keep, change, or reject it, and keep them in control without asking them to manage anything?
• Improvements that travel: What Bonsai learns from one person’s work can help their team too. How do you share it without leaking private context, and keep it improving as others use it?

What you’ll do

• Build the self improvement loop end to end: capture evidence from real work, propose improvements, evaluate them, and make the ones that help available for future work.
• Build the skill runtime: discovery, retrieval, loading, and composition inside the agent loop.
• Build the pipeline that turns session history into new skills, new memories, and improvements to existing ones, and the review surfaces where a person accepts, edits, or rejects them.
• Design and build the experience for non-technical people: what Bonsai learned, why it suggests it, and how to review, change, or undo it.
• Build evals for improvements: task suites, graders, and production signals that tell us whether a change helped.
• Own sharing: team and marketplace distribution, versioning, and provenance.

Who you are

• You’ve built with LLMs past the prompt: retrieval, evals, agent loops. You know a prompt change is a code change, and you test it like one.
• You write very well. A skill is instructions for a model, and most bad skills are bad writing.
• You care about measurement. You can design an eval that tells you something true, and you distrust one that only ever goes up.
• You have good product instincts. You know when to interrupt someone and when to stay quiet, and you can make a complex system useful to someone who has never heard of a context window or an eval.
• You lean on skills, rules files, or custom instructions heavily in your own work, and you have opinions about what makes them work.

Bonus

• You’ve built recommendation, ranking, or retrieval systems in production.
• You’ve built LLM eval infrastructure.

About Yaklabs

Yaklabs is building a desktop AI application that turns knowledge workers into AI power users, without docs, API keys, or knowing what to ask for. One surface for agentic AI work: proactive, personalized, progressively powerful. The gap between AI power users and everyone else isn’t talent, it’s setup. We close it.

The team: Alex Goddijn spent four years at Palantir, then founded an agent harness company that was acquired by Poolside AI. Seb Goddijn led internal AI at Ramp. We’re building Bonsai so that kind of leverage isn’t limited to companies with their own internal AI team.

We have funding. Product and design work in person in New York City, and some engineering roles are remote. Small team, high intensity, no passengers.

The deal

• Early team. $215k to $325k base salary for US-based hires, plus meaningful equity.
• In person in New York City for product and design roles. Some engineering roles are remote, and each posting lists its location.
• You’ll shape the product, the culture, and the company. This is a founding-team role.