Hover over the peaks to read the traceTap the peaks to read the trace
One question connects my two careers, from a mass spectrum to a language model’s answer: can this output be trusted, and by whom?
I am an MSc student in Human-Centred AI at Newcastle University, funded by a Spärck AI Scholarship, and I am looking for PhD opportunities starting in 2027. I worked with analytical measurement for about ten years, in research and industry, then spent three years building production software, including evaluating LLM systems before release. My current project is an independent audit of the health advice that language models give. If you work on any of the threads below, I would be glad to hear from you.
Four threads that run through my chemistry, engineering and AI work, all about the gap between what a system outputs and what a person can safely do with it.
Areas: AI in health and digital health · ethics of AI in health · health informatics · biomedical image analysis · fairness, bias and trustworthy AI · human–computer and human–AI interaction · chronic and under-diagnosed conditions · mental health
Spectra, images, sensor streams and health records are messy in similar ways: interference, drift and missing context. I am interested in machine learning that recovers real signal from them, from analytical chemistry to biomedical imaging.
A model, like an instrument, should show what it measured, what it is inferring and what it cannot support, so that people trust it appropriately: neither too much nor too little. This is what trustworthy AI means to me in practice: systems whose limits are visible to the people relying on them.
I want to find out whether AI is least reliable for the people with most at stake: women, neurodivergent people and people with chronic or contested conditions. My audit of language models that give health advice is a first step. I am also interested in how systems perceive and respond to behaviour, expression or symptoms that differ from the assumed norm, in mental health and beyond.
Patients, clinicians and analysts live with a system's output. I want to evaluate and design with them, especially people whose capacity changes from day to day, and to ask what values a system encodes: whose experience it treats as typical, when it should defer to a clinician, and whether it supports wellbeing rather than engagement.
An independent study, designed alongside my MSc.
People with hypermobile Ehlers–Danlos syndrome often wait years for a diagnosis, and the physical symptoms of autistic or ADHD women are often attributed to their neurodivergence. I want to know whether general-purpose language models reproduce this pattern when people turn to them for health advice.
I am designing a counterfactual audit in which the same symptom vignettes go to five LLMs while one factor changes at a time: gender, stated neurodivergence, symptom presentation, and patient versus clinical framing. A factorial design lets me attribute a change in the answer to a specific factor rather than to incidental wording. Responses will be coded with an LLM-assisted scheme only after human coders validate it (Cohen’s κ ≥ 0.70), because an unvalidated LLM judge would bring its own bias into a study of bias. Because responses are nested within models and vignettes, I will analyse them with mixed-effects logistic regression, and I plan to pre-register the design on OSF before collecting data.
Vignettes trade realism for control. If the audit finds a pattern, the next question is whether it holds for descriptions written by patients themselves, a study I would want to design with the people affected.
In every project, the hardest decision was not how to run a method but what the evidence could support. I want to bring that judgement to language models, where fluent answers make unsupported claims harder to see.
Spärck AI Scholarship. Elected course representative and member of the School of Computing's Student–Staff Committee.
Designed and built the company’s B2B platform from scratch as its sole engineer, leading a second developer who joined in the final month. The platform connects fashion brands with garment factories and tracks orders through more than 20 production stages, defined with the co-founders from their knowledge of factory workflows.
Testing with real users showed that order statuses which were logical from a software perspective did not reflect how factories actually work, and that many factory users had limited digital experience. I remodelled the workflow and simplified the interface accordingly, and verified core features with automated tests (116 for the order-reporting module alone).
Designed and built a multi-tenant retrieval-augmented generation (RAG) knowledge platform, used mainly for customer-service chatbots that answer from each organisation’s own documents with source citations.
Evaluated it before release: built golden sets of test questions with expected answers and scored responses against a rubric for correctness, tone and completeness. Checked retrieval and generation separately, testing whether the system returned the right passages and whether answers stayed faithful to them rather than hallucinating, so that each failure could be traced to its source.
Tested robustness with adversarial and jailbreak prompts, compared models and prompt versions side by side, and re-ran the test suite as regression checks after every change, alongside user feedback from production. Documented failures and argued for changes from that evidence.
Also built agent workflows using the Model Context Protocol and added natural-language interfaces to products (React, Next.js, Python/FastAPI), working closely with designers.
A golden set can only test the questions its authors thought to ask, in the words they would use; my audit addresses this by varying how the same symptoms are described.
Designed and built dashboards presenting predictive-model outputs (equipment health, bottlenecks, controller performance) on digital-twin platforms for heavy industry, for clients including Shell and Petrobras.
When feedback showed that users struggled to understand the model outputs, I revised chart types, colour encoding and terminology to make predictions interpretable for non-specialists. Designed interfaces to WCAG accessibility guidelines and added automated interface tests.
Built web applications from design to production (React, Next.js) and advised clients on accessibility. Before that, frontend developer at The Brooklyn Brothers (2023), on Unilever campaigns in more than 15 markets.
GPA 9.4/10.
Developed around 20 new colour recipes a month for textile printing, combining reflectance spectrophotometry and recipe-prediction software with iterative lab trials to match targets within tight colour-difference tolerances.
When production batches drifted from the approved standard, diagnosed the cause and reformulated a correction. This was harder than the original development, because the new colour had to be built over the off-shade fabric rather than from a clean base.
GPA 9.5/10.
Undergraduate dissertation at FURB, supervised by Dr Martinho Rau; graded 9.8/10. I sampled 85 banknotes from seven areas of Blumenau over six months to test whether Oktoberfest increased contamination. Using gas chromatography–mass spectrometry (GC–MS) with the faster of two temperature programmes, I identified cocaine by its characteristic fragment ions and NIST library matching on 76 of 85 notes (89%; 67–100% by area). The Oktoberfest hypothesis was not supported. The planned clean-up retained the cocaine itself and left the lipid interferents in the extract, and without a reference standard I could not check extraction efficiency. Rather than report quantities the data could not support, I restricted the study to qualitative identification and treated the co-extracted compounds as data: correlations between cocaine and the major peaks grouped the areas into three chemical profiles, suggesting at least three distinct sources, a hypothesis rather than a conclusion. With a certified reference standard, the next step would be validated quantification. The story behind it
Cremer (2016–2017): quality control of hospital products such as dressings and adhesive tapes. Investigated customer complaints by testing whether reported defects could be reproduced, and fed confirmed causes back to production for correction.
Umwelt (2015–2016): microbiological and chemical analysis of water and effluent for compliance with Brazilian legal discharge limits (CONAMA).
Brewing and malting school (2016): laboratory technician, preparing chemistry and microbiology labs and supporting teaching.
Independent summer research project at De Montfort University, supervised by Prof Richard Brawn. As the sole researcher, I designed and carried out a study of whether nine visually similar red-to-pink lipsticks from five brands could be told apart as trace evidence, using three low-cost techniques: ATR-FTIR (infrared spectroscopy), a video spectral comparator (VSC) and CIELAB colour measurement from photographs. Infrared peak areas varied by up to about 40% between triplicates, so I recommended absorbance intensity as the more repeatable measure. The VSC distinguished all nine products under UV, but only by combining several wavelengths, and uneven smears limited the precision of the colour measurements.
A year in the UK, fully funded by the Science without Borders Scholarship.
Undergraduate research fellowship at FURB, supervised by Prof Jürgen Andreaus; PIBITI scholarship, funded by CNPq and FURB. I led the polyester strand of a project testing whether two enzymes, cutinase and lipase AY, could replace conventional sodium hydroxide (NaOH) treatment as a milder way to make synthetic fabrics more water-absorbent and dyeable. I designed the protocols and ran the treatments; a second fellow joined near the end to take over the polyamide strand when I left for the UK. Because no single measure captures the effect, I used three: wicking tests, reflectance spectrophotometry after dyeing, and UV–Vis scans of the treatment liquors. Water absorbency rose, most clearly with NaOH and cutinase (the lipase AY results were inconclusive because of high variability), but dye uptake improved only slightly: on polyester, the gain in absorbency did not translate into dyeability. An unexpected UV–Vis signal was reported as an open question rather than as evidence of the expected product. In enzyme assays in ionic-liquid mixtures, the liquid’s own colour raised the absorbance, so every reading was corrected against its value at time zero, and the resulting negative values in some runs were reported as a limitation rather than discarded. I am first author of the project’s final report; the work was published as two conference abstracts and presented orally at MIPE, FURB (2014).
What I can bring to a lab on day one.
Systems I designed and built, and where they ended up.
Retrieval-augmented generation platform, 2025 to present
My role I designed and built the whole system on my own at Straas.tech: the RAG pipeline, semantic search, multi-tenant architecture and the evaluation process described in my research experience.
I wanted the system to be honest about the limits of its sources, so it answers only from an organisation’s own documents, cites the passages it used, and says it does not know when those documents do not cover a question. Evaluating retrieval and generation separately made it possible to check that this behaviour held.
It became one of the company’s products, used mainly for customer-service chatbots. I have continued developing my version independently as Noesis.
B2B platform for the fashion supply chain, 2026
My role Co-founder and CTO. Designed and built the platform from scratch as its sole engineer, leading a second developer who joined in the final month
Connects fashion brands with garment factories and tracks orders through more than 20 production stages, defined with the co-founders from their knowledge of factory workflows, with role-based access control and data deletion compliant with Brazil’s data protection law (LGPD) through anonymisation.
Testing with real users showed that statuses which were logical from a software perspective did not reflect how factories actually work, and that many factory users had limited digital experience. I remodelled the workflow and simplified the interface accordingly, and verified core features with automated tests (116 for the order-reporting module alone). The platform is still in operation. The lesson I took from it: a system can be internally consistent and still wrong for the people who use it.
Concept and prototype, 2026, not in active development. Best Pitch Award among 10 teams, Spärck AI Scholarship launch hackathon
My role The original idea and the working prototype; the team, scholars from Imperial, Oxford, UCL and Bristol, developed the pitch together
An agent to help patients and family members, the people who live with symptoms day to day, keep track of and report symptoms, designed for the NHS context.
The prototype left two questions I would like to study: how far a system should rely on a family member’s report when the patient cannot report for themselves, and when the agent should stop collecting information and hand the case to the clinical team.
Online group therapy and peer-support platform, 2023 to 2026. Selected for a pre-incubation programme; paused in 2026 so I could focus on research
My role Co-founder and product architect; the project began as my personal initiative
Addressed the lack of accessible mental-health support for neurodivergent and chronically ill adults. I designed it around accessibility, safety and dignity rather than engagement: professional group therapy combined with peer-support communities, a low-cognitive-load interface, and gentle progress tracking that supports continuity without pressure.
I developed it to prototype, and it was pitched at Techstars Startup Weekend Blumenau (2025). Designing it raised the questions that led me to responsible AI: what a system should surface, how it should handle sensitive moments, and when it should defer to a human.
Career platform for nurses, 2025. 1st place at the Conecta Saúde Health Tech Hackathon
My role Design lead and prototype builder, working closely with two nurses and a business specialist
A working prototype built in three days, with microlearning, career planning and one-to-one scheduling, addressing the lack of structured career progression for nurses and nursing technicians.
Healthcare management system, 2025. Final project, Tecnólogo in Systems Analysis and Development
My role Designed and built end to end, in under six weeks
A system for patient records, appointments, prescriptions and exams, with role-based access so that each clinical role sees only the information it needs.
Conference abstracts from my fellowship at FURB, published under my former name, Jennifer da Silva. I carried out the polyester work, including the treatments, the hydrophilicity and dyeing analyses and the enzyme-activity tests in ionic liquids, and presented the work orally at the conference. Kathleen Balestrin, first author of the abstracts, took over the project when I left for the UK exchange; I am first author of the project’s final report. The results only partly supported the project’s premise: enzyme treatments increased water absorption on both fabrics, but on polyester they barely improved dyeing, while on polyamide cutinase gave the best dye uptake.
Essays on rigour, AI and the people on the receiving end of both.
Read the essay“Being determined doesn't mean forcing the result you want. It means refusing to claim one the evidence hasn't earned.”
I am happy to talk about PhD projects, collaborations on LLM evaluation, or anything on this page. My full name is Jennifer Da Silva Baumgart. My university address ends after the MSc, so my personal address is the one to keep.