Junior Data Scientist

About the role

Junior Data Scientist

Phase IV | Pharmaceutical Intelligence Fully remote, United States $100,000 to $120,000 base, depending on experience Full-time | Start date: as soon as we find the right person

About Phase IV

Patients write about their medicines in public, at length, and in more detail than they give anyone in a clinic. They describe the side effect that made them stop taking something, the twelve weeks it took before anything changed, the switch an insurer forced on them, the symptom they never mentioned to their prescriber because it sounded trivial. Most of it is written in forums and on drug review platforms, in a dozen languages, and almost none of it reaches the company that made the drug.

Phase IV turns that material into structured evidence. We collect first-hand patient accounts of medicines at scale, extract what is actually being described, normalize it against the standard vocabularies the industry runs on, and make it queryable for commercial, medical affairs, HEOR and safety teams.

Two things define the work. The first is the insistence on first-hand accounts. A great deal of what is written about a drug online is promotional or secondhand: someone relaying something they read elsewhere. Separating the two is a hard NLP problem and it is the core of what we sell, because an account of what someone experienced is worth something and an account of what they heard is not.

The second is that we sit upstream of pharmacovigilance case-management systems such as Oracle Argus and Veeva Vault Safety. We are not a replacement for them and we do not do regulatory reporting. We are the layer that tells a company what patients are saying before it shows up anywhere else.

We are early and deliberately small. The corpus already runs into millions of posts and comments across the major condition forums and drug review platforms, with named-entity and relation models trained on a purpose-built annotation schema.

The role

This is a first or second job in data science, and the point of it is range. You will work across the collection and cleaning of patient-authored text, the annotation that produces our training data, the normalization of drugs and events against RxNorm and MedDRA, and the evaluation that tells us whether a model is good enough to put in front of a client.

Producing the code is no longer the difficult part of this job, and we are not hiring on typing speed. A model will draft much of what you write and we expect you to let it. The difficult part is deciding what should be built, recognizing when what came back is subtly wrong, and being able to say why one approach was right and the alternatives were not. In this domain that matters more than most: a label that is nearly right produces a finding that is confidently false about a real medicine.

You will be working directly with the founding team, including a technical founder who holds a PhD in machine learning and has published peer-reviewed research on this method. There is no layer between you and the decisions. That is the main reason to take a job like this one.

What you will actually do

Work out how a piece of data should be collected, cleaned or linked before anything is written, and be able to justify the choice

Build and maintain pipelines in Python that gather patient-authored text from forums and drug review platforms, including sources that do not want to be gathered from

Annotate text against our entity and relation schema, contribute to the gold set, and measure agreement honestly rather than flatteringly

Review machine-generated annotation, find where it fails systematically, and feed that back into the guidelines

Normalize drug mentions against RxNorm and event mentions against MedDRA, and investigate the cases that will not map

Handle the awkward parts of real patient language: negation, hypotheticals, dosage written six different ways, and accounts that turn out to be about someone else

Query and manage data in PostgreSQL, and reason about what a query is actually doing at scale

Build and run the checks that tell us whether a pipeline is quietly failing

Analyze model outputs, look for patterns worth reporting, and say so when a pattern does not hold up

Document what you built, and why you built it that way, well enough that someone else can pick it up

What we need from you

Essential:

A sound grasp of the fundamentals: statistics, sampling and bias, evaluation design, how a relational database executes what you asked it for, and what a model is doing when it returns a number

The ability to reason from a problem to an approach, weigh it against the alternatives, and defend the choice when someone questions it

A critical reading eye. You should be able to look at code, a label set or an analysis and find what is wrong with it

Working Python and SQL. You need to be fluent enough to work unaided and to judge whether output is correct, which is a lower bar than it used to be and a different one

Patience with detailed, careful work. Annotation quality decides model quality, and there is no way around that

The ability to work through an ambiguous problem without waiting to be told what to do next

Clear written and spoken English, because you will be explaining your work to people who are not data scientists

A degree in a quantitative, computational, scientific or life-sciences subject, or equivalent demonstrable ability

Useful, and worth mentioning if you have it:

Any exposure to NLP, text classification, named-entity recognition, or transformer models

Web scraping, APIs, or reverse-engineering how a site actually serves its data

Reading knowledge of a language other than English, particularly German, French, Spanish, Portuguese or Japanese

Any background in pharmacy, pharmacology, life sciences or clinical research

Familiarity with medical vocabularies such as RxNorm, MedDRA, ICD-10 or SNOMED

AWS

We are not expecting all of the above. Sound judgment, curiosity and a habit of checking things will get you further with us than a long list.

On AI tools

We use them constantly and we expect you to. There is no exercise here where you write code unaided, and no credit for doing by hand what a model does in seconds. A significant part of our annotation pipeline is model-driven.

The condition is that you understand and can defend everything you put your name to. If you cannot explain why an approach was taken over the alternatives, or walk through what a piece of code does and why it is correct, it is not finished, whoever or whatever produced it. Plausible output and correct output look identical until somebody checks, and checking is most of the job now.

How we work

Remote, asynchronous, and light on meetings. We care what you ship and how well it holds up, not when you were at your desk. We review work seriously and give direct feedback. You will be expected to do the same in return.

What you get

$100,000 to $120,000 base salary, set by experience rather than by negotiation

Health, dental and vision coverage, and a 401(k)

Paid time off with a floor of 20 days plus federal holidays, and an expectation that you take them

An annual budget for training, conferences and equipment, chosen by you

Direct mentorship from the founding team, and a realistic path into owning a modeling area within 18 months

Deep, genuinely transferable expertise in biomedical NLP, which is a narrow field and a well-paid one

The chance to join a company at the point where what you build still shapes what it becomes

Applying

Please apply directly via LinkedIn.