profile
data scientist / ai engineer with over ten years designing and shipping production ai systems, specialising in nlp and large language models at scale.
work spans bleeding-edge research, streaming data pipelines processing millions of records a week, and retrieval-augmented llm applications, delivering seven-figure annual savings across the insurance and software sectors.
sets technical direction, defines ml practice, and mentors teams toward high-impact delivery.
experience
- current role
- leading design of graph-based fraud-detection tooling for an internal investigations unit, including a normalised data model and an interactive case-review platform
- improved reliability of an internal llm-powered assistant by redesigning prompts and adding guardrails against hallucination
- architected llm pipelines analysing claim call transcripts at scale to surface early risk signals, adopted as the standard approach across the wider ml team
- built real-time nlp pipelines processing millions of records weekly to extract structured data from unstructured text, saving well into seven figures annually
- shipped ner and gradient-boosted models to support liability assessment on loss-investigation calls
- mentored junior data scientists on model development and mlops practice, introducing review standards adopted team-wide
- owned architecture and technical direction of a core text-analysis platform used across the product
- replaced a licensed third-party tool with an in-house topic-modelling and clustering microservice
- built a continuous model-evaluation framework that became the standard gate for all releases
- automated network configuration and monitoring across international trading environments to protect uptime and regulatory compliance
capability
machine learning & ai
nlp, llms, rag, langchain / langgraph, prompt engineering, supervised & unsupervised learning, clustering, transformers, model evaluation & interpretability
software engineering
python, java (spring), fastapi / flask, api design, sql, dbt, kafka, rabbitmq, graph databases, git
cloud & mlops
aws (sagemaker, bedrock, ec2), azure, docker, kubernetes, ci/cd, spark, model observability
practice
experiment design, data storytelling, mentoring, stakeholder communication
education
thesis: sentiment mining from social media content using topic model ensembles. five peer-reviewed publications; research informed subsequent production nlp system design.
graduated 2:1. dissertation on digital forensic sentiment analysis for cyberbullying detection.
publications
blair, s. j. (2021). sentiment mining from social media content using topic model ensembles [doctoral dissertation, ulster university]. https://pure.ulster.ac.uk/en/studenttheses/sentiment-mining-from-social-media-content-using-topic-model-ense/
blair, s. j., bi, y., & mulvenna, m. d. (2020). aggregated topic models for increasing social media topic coherence. applied intelligence, 50(1), 138–156. https://doi.org/10.1007/s10489-019-01438-z
blair, s. j., bi, y., & mulvenna, m. d. (2017). unsupervised sentiment classification: a hybrid sentiment-topic model approach. in proceedings of the ieee international conference on tools with artificial intelligence (ictai 2017) (pp. 453-460). ieee. https://doi.org/10.1109/ICTAI.2017.00076
blair, s. j., bi, y., & mulvenna, m. d. (2016). increasing topic coherence by aggregating topic models. in knowledge science, engineering and management (ksem 2016) (pp. 69-81). springer. https://doi.org/10.1007/978-3-319-47650-6_6
blair, s., bi, y., & mulvenna, m. (2016). sentiment classification of social media content with features generated using topic models. in proceedings of the eighth european starting ai researcher symposium (stairs 2016) (vol. 284, pp. 155–165). ios press. https://doi.org/10.3233/978-1-61499-682-8-155
sufferance
frontend
- react — not a framework, a lifestyle commitment nobody consented to.
- vue.js — react's more agreeable sibling. says it's not like react but then does exactly what react does.
- jquery — timeless, just like tar pits.
- bootstrap — the reason every site from 2014 looks like every other site from 2014.
- node.js — javascript, but now it's the backend's problem too. because apparently one place wasn't enough.
- npm — turn a 14-line script into an archaelogical expedition.
proprietary ml / bi
- sas — a programming language with the ergonomics of a tax form and pricing of a defense contractor.
- alteryx — drag-and-drop, drag your dignity and drop it at the third nested workflow.
- spss — statistics, styled like 1998, licensed like 2026.
- tableau — makes beautiful charts out of questions nobody asked.
- power bi — excel, cosplaying as a bi tool, while excel watches from the corner wondering what went wrong.
- rapidminer — drag rectangles around until someone calls it an enterprise solution.