- The website and member product
The core of the job is building the platform for the test: the website where someone learns what Helios is and orders a kit, and the member product on web and mobile where they connect their wearables and records and read what their results mean. That includes the report, the product's moment of truth, which has to be legible, honest about uncertainty, and beautiful enough that people screenshot it. You build the LLM agents that turn model outputs into language a person can act on, and the evaluation harnesses that keep them honest, because a fabricated number in a health report is the one failure we cannot tolerate.
You make all of it something people want to open every week rather than a lab report they dread.
Behind the product sits the backend: the APIs, services, and integrations that take a member from checkout through lab-partner ingestion to results delivery. You design it, you build it, and you keep it up, with monitoring, alerting, and observability deep enough to chase a problem from a model output all the way down to a frontend bug, because members make health decisions based on what we show them.
We hold sensitive health data belonging to real people, so security is a design constraint from the first commit rather than a compliance exercise bolted on at the end. Encryption, access control, audit logging, and secrets management all run through you, our vendors get held to the same standard, and the whole system operates at HIPAA-grade requirements. When something goes wrong anyway, you run the response, ideally on an architecture you designed so that no single failure can expose very much.
Research code has to become production inference, and what serves a thousand members today needs to serve a hundred thousand without a rewrite. Latency and reliability will live in your head constantly, and so will cost, since inference bills have a way of growing faster than anyone planned for.
- Data and training infrastructure
You own the machinery that makes the models possible: the pipelines from lab-partner ingestion through QC and normalization, datasets versioned carefully enough that a result from six months ago can still be reproduced, and the training and experiment infrastructure that keeps all of it maintainable as new data sources come online, wearables and records first, our own genomic and metabolomic assays later.
And then the models themselves, trained on data from our biomarker panel. Validation deserves special mention, because when a model's output feeds someone's health decisions, leakage and miscalibration stop being academic concerns. As new data sources come online, the new models are yours to train as well.