Research Associate (AI for Science)

What matters most

• Reusable scientific data deposits

  Source: Build ingestion pipelines, metadata schemas, provenance tracking and format normalization for electron microscopy, X-ray and light microscopy data so deposits work across labs. Preparation: Be ready to explain how you would preserve acquisition context and trace transformations.

• Modular model interfaces and baselines

  Source: Develop Python APIs for swapping and composing representation models, tokenizers and downstream models; collaborate on baseline tokenization and cartography pipelines others can adapt. Preparation: Think through interface boundaries and how a reference pipeline stays adaptable.

• Navigation of Gym resources

  Source: Create LLM- and RAG-assisted interfaces that retrieve from the Gym’s datasets, models, workflows and deposited challenges. Preparation: Consider how users could inspect retrieved results and identify irrelevant matches.

• Reproducible deployment across compute scales

  Source: Make components containerized, reproducible and scheduler-friendly, from single-GPU teaching instances to multi-node NUS HPC. Preparation: Be ready to discuss what must remain consistent across those environments.

• Software others can adopt

  Source: The role includes onboarding users and helping instructors build teaching-material workflows; it also calls for tested, documented software and communication with non-engineers. Preparation: Consider how you would make a workflow understandable and usable by someone who did not build it.

Likely challenges to think through

• Interoperability without losing scientific context

  Preparation hypothesis: Normalizing varied imaging formats may make deposits easier to reuse but could obscure instrument-specific metadata or artefacts. Prepare to discuss how you would validate conversions and retain provenance.

• One workflow across very different compute settings

  Preparation hypothesis: A workflow convenient on a single GPU may need different scheduling and resource handling on multi-node HPC. Prepare to explain how you would separate portable workflow logic from deployment-specific configuration and test reproducibility.

• Helpful retrieval without misleading researchers

  Preparation hypothesis: Retrieval over datasets, models and workflows can speed discovery, but a plausible result may not suit a researcher’s task. Prepare to discuss how an interface could expose source context and limitations rather than treating retrieved content as authoritative.

Topics to prepare

• Scientific imaging conventions

  Compare the metadata and common artefacts of electron microscopy, X-ray or synchrotron, and light microscopy images. Exercise: list what must be preserved when normalizing each into a reusable dataset. The posting names these modalities and asks for deposits that work across labs.

• Provenance and FAIR data practice

  Exercise: sketch a deposit schema linking raw files, acquisition metadata, transformations, and derived outputs; identify how a later user would trace each result to its source. The role explicitly covers metadata schemas, provenance, and cross-lab reuse.

• Representation and model composition

  Exercise: draw the inputs and outputs of a tokenizer, representation model, and downstream model, then identify the interface boundaries needed to swap each component. The posting calls for modular Python APIs and adaptable baseline tokenization and cartography pipelines; do not assume a particular definition of cartography without the research team's context.

• Retrieval over technical resources

  Exercise: outline how embeddings and a vector store might retrieve relevant datasets, models, workflows, and deposited challenges; name two ways a plausible-looking answer could cite the wrong resource. These are the stated targets and failure concerns for the LLM- and RAG-assisted interfaces.

• Reproducibility and HPC scaling

  Exercise: trace what changes when a reproducible workflow moves from one GPU to a scheduler-managed, multi-node run, including data access, environment consistency, and job boundaries. The posting requires scheduler-friendly components spanning single-GPU teaching instances and NUS HPC.

Skills to revise

• Build an imaging ingestion pipeline

  Using a small sample or synthetic HDF5, TIFF/OME-TIFF, or Zarr dataset, write Python code that reads files, validates key metadata, normalizes an output, and records provenance. This rehearses the explicitly stated end-to-end ingestion work.

• Implement swappable ML components

  Create a small, documented Python interface and a PyTorch example in which a tokenizer or representation model can be exchanged without changing the downstream pipeline. PyTorch is preferred, and reusable model interfaces are a core deliverable.

• Prototype grounded resource retrieval

  Index a tiny mock catalogue of datasets and workflows, retrieve entries for a question, and display the supporting entries alongside the answer; test an ambiguous query for incorrect retrieval. This mirrors the requested RAG-assisted navigation without presuming access to Gym resources.

• Package and schedule a reproducible run

  Prepare a containerized sample workflow using Docker or Apptainer and draft a Slurm job script with explicit inputs, outputs, and GPU needs. These tools are named in the posting, which emphasizes reproducible, scheduler-friendly deployment.

• Test and document for adoption

  Add tests for malformed metadata and component interchangeability, then write a short setup guide that another user could follow. The role asks for tested, documented software that people other than its author can use.

• Explain a workflow to non-engineers

  Prepare a brief notebook walkthrough of a sample dataset-to-result workflow, including what a scientist would need to supply and check. This is a preparation suggestion grounded in the stated onboarding, instructor support, interactive-tool, and communication duties.

• Map evidence to the qualifications

  Prepare a one-page index of examples covering the stated degree or equivalent-experience route, Python and an ML framework, end-to-end data pipelines, GPU work, and software others have adopted. Mark gaps as discussion points rather than claiming experience you cannot demonstrate.

• Walk through a scientific-data pipeline

  Prepare one project walkthrough showing ingestion, format normalization, metadata, provenance, and validation. Relate your design choices to the role’s electron microscopy, X-ray, or light microscopy datasets; use a proposed design if you lack a directly relevant project.

• Sketch interchangeable model interfaces

  Bring a small Python API sketch showing how a representation model, tokenizer, and downstream model could be swapped without changing the surrounding pipeline. Include proposed tests and documentation for someone who did not write it.

• Outline an adaptable baseline workflow

  Prepare a diagram for a baseline tokenization and cartography pipeline, identifying inputs, outputs, reproducibility controls, and decisions to make with ML researchers and domain scientists. Treat the diagram as a design exercise, not a claim about the Gym’s existing implementation.

• Design a retrieval-assisted navigation example

  Sketch how an LLM- or RAG-assisted interface could retrieve from datasets, models, workflows, and deposited challenges. Prepare a few test queries and explain how you would check retrieval quality and handle misleading results.

• Plan the path from one GPU to HPC

  Draw a deployment outline for a reproducible, containerized workflow that starts on a single GPU and can run through a scheduler on multi-node HPC. Identify what you would test before scaling; Slurm and Docker or Apptainer are relevant technologies named in the posting.

• Prepare an onboarding explanation

  Draft a short, non-technical explanation of how a scientist could deposit data and reuse a workflow, plus questions you would ask to understand their needs. Include a proposed way to help an instructor turn their sources and Gym content into teaching material.

• Ask about implementation priorities

  Prepare questions about the first datasets and imaging conventions, expected boundaries between data and model interfaces, how baseline pipelines will be developed with researchers, and what single-GPU and HPC environments are available.