Applied AI Engineer

## 1. Role Summary

Oncoshot turns unstructured clinical documents (clinical notes, pathology, radiology and genomic

reports) into a structured, research-grade record of each institution’s cancer population:

diagnoses, histology, TNM and overall stage, biomarkers, treatments and disease status. This

extraction runs on open-weight large language models deployed inside each hospital’s own

environment, because patient-derived content cannot leave the institution. The Applied AI Engineer

(LLM Extraction) designs, builds and improves these extraction pipelines, and is accountable for the

measured accuracy of the data they produce.

This is a hands-on engineering role, not a research position: turning oncology concepts into

precise extraction tasks, engineering prompts, schemas and pipelines around models we host

ourselves, and proving every change with numbers rather than examples. To see how we measure

extraction quality, look at our open-source evaluation framework, llmvalidate

(https://github.com/Oncoshot/llm-validation-framework), which this role works with and extends.

We are hiring at senior level by preference, and we welcome mid-level and early-career candidates

who show strong fundamentals and want to grow into the role. We invest in training, including the

clinical and oncology domain knowledge. The level and salary are set at offer, based on assessment.

## 2. Key Responsibilities

Own extraction pipelines end to end:

For each clinical document type, from classification and

chunking through structured extraction and normalisation to clinical coding standards, with a

confidence score and source justification for every extracted field.

Engineer for open-weight models:

Select, benchmark, serve and optimise LLMs that run on

hardware the hospital controls, balancing accuracy against latency and GPU cost. Decide where a

rule, lookup or curated vocabulary does the job better than a model call.

Let the numbers decide:

Evaluate every change against gold-standard datasets — precision,

recall, F1 and F2 per field, document type and institution — using the methodology defined with

the data science team, and use error analysis to choose what ships.

Own the evaluation harness and release gates:

Build the evaluation code and the regression

gates that stop a model, prompt or rule change from shipping when quality drops below the agreed

bar.

Onboard new institutions and document types:

Adapt the pipeline to how a new hospital’s

documents differ, and validate against agreed thresholds before go-live.

Ship production-grade code:

Well-tested, typed Python services and APIs, delivered through code

review and CI and supported in production.

Protect patient data:

Keep patient-derived data inside the institution and identifiers out of

logs and tickets, use hosted frontier models only for offline benchmarking on synthetic or

non-sensitive data, and contribute to AI governance and ISO 27001 evidence.

## 3. Preferred Qualifications

We don’t expect any one candidate to have all of these; strong Python, measurement rigour and

curiosity about clinical data matter most.

Bachelor’s or Master’s degree in Computer Science, Machine Learning, Data Science, Computational

Linguistics, Biomedical Informatics, or a related field.

Experience building LLM or NLP systems that run in production, with strong hands-on Python.

LLM application engineering: prompt and schema design, structured output, and multi-step and

long-document pipelines.

Serving and optimising open-weight models (e.g. Hugging Face, vLLM), including quantisation and

GPU cost/latency trade-offs.

Evaluation literacy: precision, recall, F1 and F2, error analysis, gold-standard datasets, and

regression testing of models and prompts.

Software engineering: testing, typing, API design (e.g. FastAPI), containers and CI, on cloud

infrastructure (AWS, Azure or GCP), including GPU workloads.

Experience with clinical or biomedical text is strongly preferred (e.g. pathology or genomic

reports, ICD-O-3, AJCC TNM, SNOMED CT, RxNorm, HGNC), as is familiarity with health-data privacy

frameworks (DPDP Act, PDPA, GDPR, HIPAA).

## 4. Core Competencies

Pragmatic model judgement:

Picks the simplest approach that meets the quality bar, whether a

rule, a smaller model or a better prompt, and knows when an LLM is the wrong tool.

Clinical precision:

Turns ambiguous medical concepts into exact, testable extraction

definitions, and asks a clinician when the answer is not clear-cut.

Communication across audiences:

Explains model quality, uncertainty and trade-offs credibly to

clinical, commercial and engineering stakeholders.

## 5. Success Measures (First 3 Months during Probation Period)

The scope of these goals is agreed at offer according to level.

Took ownership of at least one existing extraction pipeline: running its evaluation, diagnosing

its errors and releasing changes independently.

Published a per-field quality baseline for the pipelines owned, with thresholds agreed with the

CTO.

Delivered at least one measured accuracy improvement, or a new document type or field, through the

regression gates without a quality drop elsewhere.

## 6. Working Model

Probation period:

3 months, followed by conversion to permanent staff.

Reporting line:

Reports to the Chief Technology Officer.

Location & arrangement:

Hybrid, based in our Thane (Mumbai) office, with core working hours

overlapping Singapore, which is 2.5 hours ahead of IST.

Team structure:

Works with clinicians and the clinical validation team on what “correct”

means for each field, with the data science team on how quality is measured, and with Oncoshot’s

platform, data pipeline and product teams.

Cadence:

Agile/Scrum, with sprint planning, refinement and reviews.

## 7. Salary Range

Indicative annual CTC: ₹10,00,000 – ₹40,00,000 (₹10 – 40 LPA), based on experience and seniority.