New opportunity

Research Engineer, Applied ML

Triomics · HSR, India

About this role

Triomics seeks a Research Engineer to develop applied NLP and LLM systems for healthcare and oncology workflows. The role involves clinical text, patient-trial matching, summarization, RWE abstraction, information extraction, and structured outputs from clinical documents. It is an on-site, full-time role at the India office.

Responsibilities

Build prototypes and production-oriented components for clinical information extraction and LLM workflows. Implement prompts, RAG pipelines, fine-tuning experiments, extraction logic, evaluation scripts, and error-analysis tools.
Investigate clinical-text model failures and work with senior engineers to improve model quality, robustness, and reliability.
Collaborate with Clinical AI Data Specialists on labeling guidelines, data quality, and clinical edge cases, and with ML Evaluation Engineers on benchmark and regression tests.
Write clean, tested, maintainable Python code for internal tools, data processing, and annotation and evaluation workflows. Within six months, the successful hire is expected to own bounded workstreams, document reliable experiments, suggest improvements from error analysis, and collaborate with clinical data, evaluation, and MLOps teams.

Requirements

1.5–4 years of hands-on experience in ML, NLP, LLMs, or data-intensive software engineering; strong Python skills and willingness to work with messy real-world data.
Familiarity with PyTorch, HuggingFace, embeddings, transformers, RAG, prompt engineering, and basic fine-tuning. Understanding of precision, recall, F1, train/validation/test splits, overfitting, error analysis, and experiment tracking.
Ability to learn clinical concepts, clarify labeling and data questions, and take ownership of assigned modules.

Preferred

Experience with clinical NLP, biomedical text, healthcare documents, or information extraction; LangChain, LlamaIndex, vector databases, spaCy, OCR/document processing, annotation tools, or structured LLM outputs; and exposure to Docker, APIs, ML deployment, MLflow, W&B, or data pipelines.

Benefits listed include office lunch, flexible working hours, family health insurance, Zomato meal benefits, and company-sponsored workations.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

PythonMachine LearningNatural Language ProcessingLarge Language ModelsClinical NLPInformation ExtractionRAGPrompt EngineeringFine-tuningPyTorchHuggingFaceEmbeddingsTransformersModel EvaluationError AnalysisExperiment TrackingLangChainLlamaIndexVector DatabasesspaCyOCRDocument ProcessingDockerAPIsML DeploymentMLflowW&BData Pipelines

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Build clinical NLP and LLM workflows
The role builds prototypes and production-oriented components for information extraction, patient-trial matching, summarization, and structured outputs. Prepare to explain how you would choose and test prompts, RAG, extraction logic, or basic fine-tuning for a defined clinical-text task.
Improve models through measured error analysis
The posting calls for evaluation scripts, benchmark and regression tests, and analysis of clinical-text failures. Prepare a clear approach to dataset splits, precision, recall, F1, experiment tracking, and turning recurring errors into testable improvements.
Resolve data and labeling questions collaboratively
The engineer works with Clinical AI Data Specialists on guidelines, data quality, and clinical edge cases, and with ML Evaluation Engineers on tests. Prepare to describe how you would document an ambiguous label, ask for clarification, and carry the decision into evaluation.
Own reliable, maintainable workstreams
The posting expects clean, tested Python and, within six months, independent ownership of bounded modules or workstreams with clearly documented experiments. Prepare to explain how you would scope a component, test it, and make its results reproducible for collaborators.

Likely challenges to think through

Ambiguous evidence in clinical documents
A likely difficulty is distinguishing absent information from unclear or conflicting text when producing structured outputs. Prepare an example-based method for inspecting source passages, recording uncertainty, and escalating label questions rather than forcing an answer.
Balancing error types by workflow
For tasks such as patient-trial matching or information extraction, false positives and false negatives may have different consequences; the posting does not specify acceptable thresholds. Prepare to explain how you would compare precision and recall by use case, review representative errors, and agree on evaluation criteria with the relevant team.
Moving from experiments to dependable components
Trying prompts, RAG, and fine-tuning can yield improvements that do not hold up on new clinical text. Prepare to discuss controlled comparisons, held-out and regression tests, and tested Python interfaces before recommending a prototype for broader use.

Practice your answers

Technical practice

How would you extract clinical facts from an unstructured oncology document?

What to coverThe role includes clinical information extraction from documents such as pathology reports and clinical notes. Suggested preparation: explain how you would define fields with clinical data specialists, handle ambiguous or conflicting text, and preserve evidence for each extracted value.

How would you approach patient-trial matching from clinical text?

What to coverPatient-trial matching is a stated workflow. Suggested preparation: describe how you would turn relevant record text into matchable information, distinguish missing evidence from a mismatch, and evaluate errors before relying on a match.

How would you distinguish clinical summarization from RWE abstraction?

What to coverThe posting names both summarization and RWE abstraction. Suggested preparation: discuss how a readable summary and a defined set of abstracted fields require different output checks, particularly for completeness and support in the source document.

How would you design a RAG pipeline for questions about clinical documents?

What to coverImplementing RAG pipelines is an advertised duty. Suggested preparation: cover document preparation, retrieval with embeddings, use of retrieved context in the prompt, and how you would detect unsupported answers or missed evidence.

When would you try prompt changes rather than a fine-tuning experiment?

What to coverThe role involves prompts and basic fine-tuning experiments, with familiarity in transformers, PyTorch, and HuggingFace. Suggested preparation: compare data needs, effort, and evaluation for each approach, including how you would check whether an apparent gain generalizes.

How would you evaluate an information-extraction model?

What to coverThe posting explicitly calls for precision, recall, F1, and train/validation/test splits. Suggested preparation: explain the consequences of false positives and false negatives for a defined extraction task, then describe how you would use separate data splits to detect overfitting.

How would you investigate a drop in clinical-text model quality?

What to coverAnalyzing model failures and running benchmark and regression tests are stated responsibilities. Suggested preparation: outline how you would inspect error examples, distinguish data or label problems from model problems, and test whether a proposed fix improves results without introducing regressions.

How would you make LLM-generated clinical fields reliably structured?

What to coverStructured outputs from clinical documents are part of the role, and structured LLM outputs are listed as a nice-to-have area. Suggested preparation: discuss a defined output format, validation of required fields and values, and what to do when the document does not support a value.

How would you build a maintainable Python workflow for messy clinical documents?

What to coverThe role calls for tested Python tools and data-processing scripts; OCR and document processing are nice-to-have areas. Suggested preparation: describe input checks, handling of malformed or incomplete records, and tests that make extraction or annotation workflows easier to maintain.

How would you document a model experiment so another engineer can assess it?

What to coverThe posting expects reliable experiments with clear documentation and knowledge of experiment tracking. Suggested preparation: specify the data split, model or prompt version, configuration, metrics, and error examples you would record, then explain how you would compare the result with a baseline.

Behavioural practice

Tell me about a time you resolved an ambiguous labeling or data definition.

What to coverThe role involves clarifying labels and clinical edge cases with Clinical AI Data Specialists. Suggested preparation: use a real situation, the questions you asked and actions you took, and the resulting decision or improvement.

Tell me about a time you used feedback to improve a failing ML component.

What to coverThe role calls for working with senior engineers to improve model quality and reliability. Suggested preparation: describe the failure, how you incorporated another engineer’s input, what you changed, and the observed result.

Tell me about a time you worked with others to evaluate a model change.

What to coverCollaboration with ML Evaluation Engineers on benchmarks and regression tests is stated work. Suggested preparation: give a real example of how you agreed on a test, handled differing interpretations of results, and decided what to do next.

Tell me about a bounded technical workstream you owned.

What to coverIndependent ownership of a module or bounded workstream is listed as a six-month success outcome. Suggested preparation: explain the scope and constraints, your actions and documentation, and the result without overstating your individual contribution.

Tell me about a time messy data changed your implementation plan.

What to coverThe posting calls for willingness to work with messy real-world data and for reliable Python code. Suggested preparation: describe the data issue, how you investigated and adapted the workflow, and what the outcome taught you about testing or reliability.

Sharpen your knowledge

Topics to prepare

Clinical-text interpretation and labeling
The posting names oncology documents, information extraction, patient-trial matching, and collaboration on labeling guidelines. For revision, inspect a few fictional clinical notes and write down which facts are explicit, which are ambiguous, and what you would ask a data specialist before assigning labels.
Transformers, embeddings, and RAG
Familiarity with these concepts is expected. Sketch how a clinical-document passage would be embedded, retrieved, and supplied to an LLM; then identify where retrieval errors could affect the answer.
Prompting and fine-tuning trade-offs
The role implements prompts and basic fine-tuning experiments. Compare, on paper, when you would try a prompt change versus a fine-tuning experiment for a recurring extraction error, and specify what evidence would justify the choice.
Evaluation design and overfitting
The posting explicitly calls for precision, recall, F1, train/validation/test splits, and experiment tracking. Work through a small extraction confusion matrix, calculate the metrics, and explain how you would use held-out examples to check a proposed improvement.
Clinical-text failure analysis
Analyzing failures and improving robustness are stated responsibilities. As preparation, group hypothetical extraction errors by ambiguous wording, missing context, and inconsistent labels; propose one testable change for each group.

Skills to revise

Tested Python data processing
The role requires strong Python and clean, tested code for messy data. Practice writing a small parser that converts inconsistent fictional clinical-note fields into a consistent record, with tests for missing fields and malformed entries.
Structured information extraction
The posting calls for extraction logic and structured outputs from clinical documents. Practice defining a small output schema, extracting explicitly stated facts from fictional notes, and recording missing or uncertain values rather than guessing.
Prompt and RAG prototyping
Implementing prompts and RAG pipelines is listed work. As a practice exercise, build a small example using fictional passages: retrieve supporting text, generate a structured answer, and inspect cases where the answer lacks support.
PyTorch and HuggingFace experimentation
Familiarity with both and basic fine-tuning is expected. Rehearse a small text-labeling experiment, keeping training and held-out examples separate and recording settings and results so the comparison is reproducible.
Evaluation scripts and regression checks
The role involves evaluation scripts, error-analysis tools, and benchmark and regression tests. Write a script that compares predicted and reference labels, reports precision, recall, and F1, and lists changed errors between two runs.

Your preparation checklist

Select a relevant project walkthrough
The role involves applied NLP and LLM work on clinical text. Prepare a walkthrough of an actual project, if available, covering the data, your contribution, model or pipeline choices, results, and limitations. An information-extraction example is especially relevant.
Sketch a clinical-document extraction design
As a practice exercise, outline how you would turn a pathology report or clinical note into structured fields. Show how you would handle ambiguous text, missing values, validation, and questions for clinical labelers; do not present the sketch as a Triomics system.
Prepare a model-choice comparison
The posting names prompts, RAG, and basic fine-tuning. Make a one-page comparison of when you would test each for a clinical-text task, including the data needed, likely failure modes, and how you would choose a baseline.
Build an evaluation and error-analysis example
Prepare a small example showing train/validation/test separation, precision, recall, F1, and a categorized error review. Include a proposed regression test for one failure, reflecting the role’s evaluation responsibilities.
Gather evidence of maintainable Python work
The role calls for tested Python code and ownership of bounded modules. Choose an actual code example, if available, and be ready to explain its tests, handling of messy data, documentation, and a change you made after feedback.
Practice a labeling and collaboration discussion
Draft a few clarifying questions about annotation guidelines, conflicting labels, and clinical edge cases. Prepare an example, if you have one, of resolving a data-quality or evaluation issue with another function.
Bring focused questions for the team
Ask how this role divides prototype and production-oriented work, how clinical data specialists and evaluation engineers review failures, and what ownership of a bounded workstream looks like in the first six months.

Availability last checked: 4 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in India. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…