justinyan

Xaira Therapeutics | Midlevel/Senior/Staff Engineers | Seattle or SF

Xaira Therapeutics is developing biological foundation models (not LLMs) to hit undruggable targets and cure disease. We've put together an incredible cross-disciplinary team of pharma veterans who have brought drugs to market, drug designers/AI scientists from Nobel laureate David Baker's lab, and engineers from tech in order to meet the moment and build a company to reimagine how drugs are made in the pursuit of better medicines. We're not a SaaS company, and the ultimate goal is to see the medicines we make delivered to patients.

I lead the engineering team and am hiring across multiple teams. There's:

- Research Engineering: work on a modeling team with the goal of making a model better. You think in tensors, understand what makes a kernel fast, and obsess over training throughput.

- X-Scientist: building an agentic copilot for drug development with access to our internal suite of tools. You have backend experience, have built tools for agents to interact with, and enjoy tinkering with agent harnesses.

- Data Engineering: we generate large volumes of data from our lab that provide critical feedback loops to our modeling efforts. You either hate or love lakehouses, have a preferred ETL orchestrator, and are intrigued by the problem of capturing data off of lab instruments.

- ML Platform: boy do we have some GPUs for you. Modeling teams need productive compute infrastructure, job scheduling, and numerous other affordances so that they can "just focus" on the model. You shudder at the words "falling off the bus", you have an opinion on how checkpoints should be recorded, and you've built tools that improve the modeling lifecycle.

Send me an email to apply: [email protected] Include a resume and tell me about something you've built that illustrates your skills.

simonw got me interested in the idea after I ran into one of his blogposts talking about how ziglang does it, and then studying ruff + maturin helped me reverse engineer how to accomplish this without any python intermediary in a "python package".

I put together a cruft template (https://github.com/justin-yan/templates-cruft/tree/main/pack...) since I had the exact same thought as you once I got it working - why not do this with a bunch of different tools? I did it with a couple of my favorites (just, lazydocker) but haven't gone back to clean things up/make things a bit more systematic. So it's a pretty messy weekend project (caveat emptor) but hopefully it's a helpful starting point if you want to try it!

I'm a fan of using just as a command-runner as well, but took a slightly different approach to making it pip-installable! I created the `pyjust` package (https://pypi.org/project/pyjust/), which literally just packages the `just` binary in platform-specific wheels, and then unpacks it onto the PATH by leveraging the `scripts` directory unpacking.

Why do this you might ask? My rationale here was that I use `pipx` to install some python CLIs, and it was nice to be able to lump `just` in with that, as well as to be able to declare it as a "dependency" and pin specific versions for distribution.

Taking a different line of questioning: Does anyone have any insights or resources on the non-tech stack a one-man SaaS should use?

At various points, I've looked into opening a single-owner LLC to hold some small side projects I want to put on the internet for fun, but the number of decisions and amount of overhead has always seemed pretty daunting.

The act itself seems to revolve around limiting section 230 protections (https://uscode.house.gov/view.xhtml?req=(title:47%20section:...) to those who pass a checklist of TBD requirements determined by a TBD bureaucracy.

I've struggled a bit to find resources helping to describe what section 230 protections are actually for, however. Many of the examples explaining section 230 protections seem to revolve around things like defamation in Youtube comments or product reviews, but I'm having trouble making the leap from that to why messengers are so concerned about this bill. Why do tools like messengers need section 230 protections to begin with? AFAICT they still have to submit to things like national security letters, so what does section 230 buy, say, Whatsapp or Signal?

The Atlantic has an effort to break this tracking down on a state-by-state basis (https://covidtracking.com/data/) which definitely has some interesting nuggets and gives some color on how they arrive at their numbers on a per-state basis.