Gather
Find candidate evidence across research and attention-tracking platforms.
02 / Bath Insights · June–July 2026
Research leaves traces. I helped find them.
A Research & Impact internship combining evidence review, a Python and local-LLM workflow, and the less visible work of making a tool useful to colleagues.
01 / The problem
Impact isn’t confined to the research paper.
A paper can be cited in policy, discussed in the media or used by organisations outside academia. During Bath Insights, I worked with another intern to help evaluate evidence for approximately four REF2029 impact case studies.
That meant looking across PURE, Altmetric, PlumX and Overton, reviewing material associated with 50+ research outputs and triaging more than 200 evidence items a day. The challenge was to turn scattered material into structured evidence that colleagues could review.
02 / What I built
From evidence discovery to a reviewable dataset.
Structure
Use a Python pipeline with Gemma 3, hosted locally through Ollama, to support classification.
Review
Export structured Excel datasets for colleagues to inspect and use.
I also developed an internal web interface and supporting documentation. The workflow assisted evidence discovery and organisation; professional judgement remained essential when assessing whether an item supported an impact claim.
03 / Working with people
The interface was only part of the job.
I presented the workflow to the Research Impact Lead and used the feedback to refine the interface, output structure and documentation. I worked directly with the eight-person Research Excellence and Impact team and contributed tools and training resources for a wider group of approximately 30 research-impact professionals.
Alongside the technical work, I supported the Research Excellence Network, contributed to REF impact case studies, and worked with colleagues across Research & Impact Services and Public & Policy Engagement. Conversations with researchers, policymakers, charities and industry partners helped me understand what “impact” looks like beyond a metric.
A workflow has to make sense to the person reviewing the evidence, not just to the person who wrote the code.
The internship gave me practice in explaining technical work, taking feedback and changing the implementation around the people who would use it.
04 / What this demonstrates
A technical idea, carried through to handover.
- Applied automation: Python, local model integration and structured output in a real research-support setting.
- Evidence judgement: working across several sources while keeping discovery, classification and interpretation distinct.
- Communication: presenting to colleagues, incorporating feedback and writing materials that support use after handover.
This page describes my contribution and the workflow. Internal evidence, research records and university systems are not reproduced here.
Experience & timeline ↗